---
title: 'AI & Data Fest 2025 by Codebasics | Mega Offline Event'
source: 'https://youtube.com/watch?v=x-3Mefy_iZ0'
video_id: 'x-3Mefy_iZ0'
date: 2026-07-31
duration_sec: 29192
---

# AI & Data Fest 2025 by Codebasics | Mega Offline Event

> Source: [AI & Data Fest 2025 by Codebasics | Mega Offline Event](https://youtube.com/watch?v=x-3Mefy_iZ0)

## Summary

The AI & Data Fest 2025 by Codebasics is a full-day offline conference recording featuring keynotes, technical talks, and panel discussions on AI, data engineering, and career growth. It captures practical insights on AI agents, high-performance data processing with FireDucks, business-focused AI, and emerging standards like Model Context Protocol (MCP).

### Key Points

- **Event scale and sponsors** [06:32] — 50+ industry experts, 750+ attendees, 60% data professionals; sponsors include FireDucks and Agno, community partner Indian Data Club.
- **Coding advice warning** [13:32] — Keynote speaker warns: if someone tells you not to learn coding because AI will do it all, that is one of the worst pieces of advice you can get.
- **HDF framework for hype** [15:17] — The HDF (High Direction) framework helps you cut through hype and misinformation; be ready to separate truth from influencer-driven claims.
- **Challenge: hallucination and compliance** [20:49] — First production challenge with LLMs is hallucination; example: a bank chatbot promised a refund not in policy, leading to a lawsuit.
- **Challenge: data quality and governance** [22:02] — An anomaly-detection project for IoT turned into a data engineering grind; new data kept degrading model performance, forcing continuous feature engineering and retraining.
- **AI jobs will grow but not fully autonomous** [25:11] — World Economic Forum says AI/ML jobs will grow; however, human oversight, deployment, and careful engineering remain essential.
- **AI agents are companions, not replacements** [01:12:53] — Priya Singh uses the example of a dosa shop owner: an AI agent providing data-backed sales insights does not replace the owner's job.
- **Anatomy of an AI agent** [01:17:38] — An AI agent = LLM + memory + knowledge + tools; adding calendar access and a knowledge base lets agents book meetings and remember context.
- **FireDucks: a faster pandas** [01:43:55] — Sarthak Saha introduces FireDucks, a pandas-compatible dataframe library with a compiler, multi-threading, and lazy execution.
- **Projection pushdown optimization** [01:48:10] — Only load and process the columns you need; this drastically reduces memory consumption and execution time for large datasets.
- **Zero-code migration to FireDucks** [01:59:21] — Use `import fireducks.pandas` or an import hook; existing pandas code runs unchanged while getting FireDucks' performance.
- **FireDucks benchmark: 141x faster** [02:11:44] — On TPC-H benchmarks, FireDucks is ~141x faster than pandas, DuckDB ~199x, and Polars ~58x.
- **Freshers get hired by proof of work** [02:43:08] — Panelists emphasize building in public, maintaining a strong portfolio, contributing on GitHub/Hugging Face, and winning hackathons rather than relying on degrees.
- **Business AI: hybrid systems win** [03:25:30] — Karan shares a production case: classifying 100k+ messages using rule-based + BERT + LLMs to balance cost, explainability, and accuracy.
- **Agent Hub: reuse AI agents** [05:38:36] — Mrinal Kumar presents Agent Hub, a centralized repository where agents can be published, discovered, and reused across frameworks like agno, CrewAI, and LangGraph.
- **Model Context Protocol standardizes LLM context** [07:12:55] — Snea explains MCP from Anthropic: LLMs connect to external data sources via MCP servers, making it easy to swap tools and build context-rich applications.
- **Data engineering playbook** [06:07:49] — Snit Alam Raju outlines three pillars: infrastructure, data, and process — covering scalability, storage choices, feature engineering, and MLOps.

### Conclusion

The AI & Data Fest 2025 delivered practical, no-nonsense advice: focus on real problems, embrace hybrid AI systems, and build in public to grow your career. The event highlighted that while hype is everywhere, grounded engineering, business alignment, and continuous learning remain the true differentiators.

## Transcript

have some key points to be known by everybody in the uh session so I have a presentation over here so yeah I can say that this is the best and the biggest AI in data first why because we have 50 plus industry
experts and 750 plus attendees across India and mostly 60% are data professionals and we also have huge n working stas your favorite LinkedIn star
where you can connect with them and you also can have a great conversation with them and that is the reason I'm saying this is one of the biggest event of this this is one of the biggest event of this year
okay so we got the Platinum sponsor who are fire ducks and agos so I thank them and we also have our community partner Indian data club we have their boots next to the engagement area so I thank them and we have our Hospital
thank them and we have our Hospital partner caran hotels I also thank them and before starting the session we need to know something very very important do you guys know which is the most important place in this whole
entire building any idea the very important place no idea restroom guys so we have next room uh restroom it's on the left side so please uh you know uh
connect with volunteers if you need any help in navigating to restroom then we encourage you guys to stairs no to walk through stairs because uh the crowd is very huge please use staircase in instead of lift
then yeah we have Cafe next to this place and the lunch area is ground floor as you guys know then we have engagement Zone where we have huge boots over there
and we have games where you can win a cool goodies over there so please go and visit those sponsors booth and we also have multiple booths over there so visit and have a conversation with them to win cool
goodies and yes so you'll be having ID card right could you please turn uh the other side okay W thank you so when you can side okay W thank you so when you can see the agenda of the uh event so you'll
be uh having all the timings of the sessions if you guys have uh you know registered for the exclusive workshops you also can go through the agenda so [Music] agenda then yeah the sole reason we are
all here is to learn right from different experts across the country so the main reason is the learning part so as we came here as a community to learn
something new and to you know interact with other fellow people so I want to with other fellow people so I want to guys be kind and respectable then uh yeah let's listen with so much attention to the speakers and also to to the
to the speakers and also to to the fellow recently met so many people in got many events and I can say really those people
are kind of meaningful connections that I've made so I encourage you guys to talk to your fellow people over there and make some good connections okay then uh if you guys who need any
kind of assistance anything please reach reach out to this number is Mr s or else you also have the wallet is next to you so they can assist you in whatever assist that you need then yeah the whole reason the
other reason is have some fun take it easy learn okay while you guys have fun guys because it's going to be a memorable day so if you guys do uh take
much pictures or post that in your LinkedIn or Instagram or any other social media platform and tag us use this hashtag guys AI data s 2025 all
this hashtag guys AI data s 2025 all set okay thank you so much set okay thank you so much yeah so
some people on the stage who actually don't need any introduction kind of do you have any guesses guys we going to have two people onto the stage any have two people onto the stage any guesses so we are going to have D and on
to the stage to give a keynot so I rather have both of them onto the stage so let's have a huge Applause for them guys guys baby baby
baby baby [Music]
hello everyone good morning [Applause] hello everyone everyone Nam how are you
all okay so we are going to kick off this keyn session with the question that this keyn session with the question that I have for you there are a lot of ads happening which one excites you the most the one that excites me the most is this
so I can give 90% ofing to the I can just this just this 10% you think you can do 10% you think you can do for yeah he says if you not learning
coding and someone gives youer advice to not learning coding it is one of the wor get okay a
might Rising your also it will help you the reality through the layers of time and misation so that you can unlock your
misation so that you can unlock your career potential in the hello everyone of
how we need to know this this is called hdf how
knows this any guess this guys disappointing this is high Direction disappointing this is high Direction framework it just got in
framework it just got in before me in the next 15 minutes this is going to change your life by the way you're going to consume information if someone talk like that that's a
h and you should be ready with your truth suing St there are many examples for example Nvidia CEOs is saying all the cing will
be done by I and you don't need to learn coding I think was this what was the coding I think was this what was the name okay that was supposed to replace all the software engineerings and Elon Baba says that he
said this by the way 10 months back this year 2025 we are going to see AI you guys know what is AI yes right so artificial general
but think all these are such a responsible people they won't say responsible people they won't say things yeah good question so how about I things yeah good question so how about I get Elon to answer the
question fortunately fortunately there are people who just don't say things are people who just don't say things like this
popular than people who speak through that's a plain reality of the world that we live is one of the three Godfathers of the plan there are other two folks anyone know
like and Joshua this guy is a father of Network right you should be listening to him he is an educator he doesn't have agenda okay he a chief scientist at meta but at his heart he's an educator who
cares about truth more than anything else so I understand right like people like uh they don't have any agenda or any pressure to PO something how many of
you have the pressure to post I see a lot of influences you influen here how something other you lose right it's like you have to say something
something talk about reality now so I am co-founder of EIC Technologies it is an co-founder of EIC Technologies it is an Services compy you will see of Services compy you will see of here of here we work on the real AI
projects okay many plans many projects and what you see is all these businesses and Enterprises they want to ride on Enterprises they want to ride on a so a typical sales score okay we look
something like this client would be like you know I have billion project uh you know I have billion project uh budget I want to do something with and budget I want to do something with and like what what you want to do not sure
just throw some random ideas and then they just start the project so many people are trying to figure out how they can ride this wave and due to that what you see in the production nowadays is
bunch of P proof of Concepts many of you will be an engus here right so how many of you are noticing the
production still experimenting you build something with Lama and something new come up so now you want to PL a new mall and see how it you want to PL a new mall and see how it behaves also in terms of your real life
behaves also in terms of your real life see you all use J Bank of America non of whatever your is and Link it you have whatever your is and Link it you have all the apps they will have
interacting with LM power at I don't us this is by
have what I see is these are still old rule based dial check I don't see going rule based dial check I don't see going into production at scale yet despite the fact that J GP has been there since years and this whole thing is kind of
should have there are so many challenges with when we onject at the first challenge that we see is
hallucination and compliance can tried releasing their LM compliance can tried releasing their LM power chat board in production then what happened is there was a customer okay there was a
customer who was talking with this lard and he had a question on bment policy so he had some with his family and like okay can I get a refund and Chad confidently said yes you are eligible
for the refund we will send an email blah blah blah the actual policy was opposite they did not have this refund canul canul so this whole thing B into a lawsuit
so this whole thing B into a lawsuit okay so although you have R enabl agent okay so although you have R enabl agent grab chat board what happens is in the grab chat board what happens is in the end these models are probabilistic it's
like probabil probabilistic model with too much knowledge way too much knowledge and when you have way too much knowledge the chance that hallucination will happen slightly okay the second Big Challenge is data
quality and governance we are working with a North America based plan and there into iot devices we are building anomal detection models we use the most state of the art feature
engineering techniques such as techniques such as time we also find to our model and model was doing good on the training data that was now how this training data was
provided someone manually curated that right so the quality was decent the performance look good then you started getting new data and when we the model on the new data again performance great so then you go into this infinite Loop
of feature engineering funing Reta R and then at some point you stop you get tired you realize you need to fix get tired you realize you need to fix your data infrastructure so this expy
your data infrastructure so this expy project that started with ai ai project project that started with ai ai project quickly turned into a boring data engineering I think these two are problems I but I think there is a bigger
problems I but I think there is a bigger problem can anyone tell what is respons yeah yes of context retention cont rtion
okay so we talking about the case is starting to get but let me tell you case where you can't even lot ofies lot of this we are still
for years you years you can't it will take even more time so let's in 3 months whatever happens 6 months when people are saying 3 months
months when people are saying 3 months of [Music] let's you
know I worked with peopleold acoss different and you know the how can so with something like aard you can
just back provide them the but just think about the espe the it's not that easy to that's where you get start that's where the you get start that's where the Enterprise adoption is
really super so think so what saying
ask so World Ecom for have you seen this fut of job
Rett much those jobs are going toow in the next so now on one now on one saying do all the things make it easier
saying do all the things make it easier and like this that says in next so so what I I this are but probably the title and
responsibilities never change you won't probably call might I see the way ofing change because now are going to do and you'll be doing more stuff more
do and you'll be doing more stuff more work this yes so as you say right job roles are evolving so if you're a software software engineering means is fundamentally changing and before we go
into how exactly is changing one part that you all will have is okay World economic forum is saying that for example AI machine learning jobs and example AI machine learning jobs and data anal jobs and big dat special jobs
are going to grow but isn't it that these jobs will be but isn't it that these jobs will be fully autonomously done by AI well why fully autonomously done by AI well why not
because where couple of f deploy this whiteing production and this hackers hackers easy just go system you want easy just go system you want so these kind of inci are now happening
going back to the evolution of J let's pre your software developer working for you go to job at 9:30 you spend your next 8 hours doing
technical world time that it takes to do that work world time that it takes to do that work is let's assume it is 4 is let's assume it is 4 hours you add expectation like okay how
hours you add expectation like okay how about
involing from professional 1.0 to professional 1.0 to professional 2. in 2.0 are profession
who you have to betic in approach take enough interest in business and also also do your technical work along with this other beautiful soft skill and you are now becoming a problem solver not
solver not just get talking about these two I have a friend the founder of Vy Technologies 700 plus software
engine and that the way is evolving in and that the way is evolving in this is previously they would have this this is previously they would have this ja
hiring software Engineers as a software engineer you should be able to goet Java engineer you should be able to goet Java tomorrow you want to use programming of you should be able to in that why because
because withing things are becoming easier
and one paradig about AI it is just a next World model there are other research is going on other research is going on as how many of you have about world
model very this what happens when you follow follow the influencers and not
theat so there has been an architecture called and is contributing the reason heing it is this point is very very clear
have it will be able to navigate the world much better way compar to this
one model is next big and also we have arure we have not
receiving Point yes it may happen in years 10 years 20 years but years 10 years 20 years but you it will take all the time and when you it will take all the time and when we see the breakr it
will already started but some people are very I Wasing there
amazing area why 15
first one is like should be able to lot of every company wants to put a name on the top so they will
embra to learn unar and don't think that
skills today we are going to see see a lot of roling around this and I hope you going to have a great time thank [Applause]
so thank you for the presentation so me as a host for a technical session is the very first time but I have been to many technical session as an ATT so when I used to sit there as an ATT so many
speakers would come on the stage and used to talk right so maximum for 10 to or 15 minutes I be active for the rest of the time I more active no
thinking that I'm the best the whole but you know what I be thinking man this it's not so interesting this is what I thought in my mind but the face reaction thought in my mind but the face reaction was just not to you know the speaker but
personally I like the whole presentation and the hum tou they have had the whole presentation is to good and it was very and
with them don't face them just go so you guys your change your seat so session uh there is a CFE so
where you can have some meaningful connections talk talk talk to the unnown random people and thank
you 30 minutes make sure to be on thank you
[Applause] [Music]
[Applause] [Music]
[Applause] [Music]
[Applause] [Music] [Music] oh
feel it's a new dawn it's a new day it's a new life for me yeah it's a day it's a new life for me yeah it's a new dawn it's a new day it's a new life new dawn it's a new day it's a new life for
[Music] me and I'm feeling good me and I'm feeling good [Music] fish in the sea you know how I
sea you know how I feel River Running Free you know how I [Music] feel Blossom on the tree you know how I feel Blossom on the tree you know how I feel it's a new D it's a new day it's a
new life for me life for me and I'm feeling
what I mean don't you know butterflies all having fun you know what I mean sleep in peace one day is done
that's what I mean and this old world is a new world in a bold world for me yeah
[Music] feel know iom is mine and I know how I feel it's a new mine and I know how I feel it's a new dawn it's a new day it's a new
dawn it's a new day it's a new life for
me I'm feeling good
it's a new dawn it's a new day it's a new for
[Music] me [Music] and fishing the sea
you know how I feel Riv running free you know how I
[Music] feel it's a new day it's a new for me and I'm feeling good
me and I'm feeling good [Music] [Music] dragon fly out in the sun you know what I mean don't you know butterflies all having fun you know
what I mean sleeping peace when day is that's what I mean and this old is a new world and a
bold world for me [Music] St shine
know know I feel freedom is feel freedom is mine and I know how I feel it's a new
it's a new day it's a new life for
I'm feeling good
it's a new dawn it's a new day it's a new life
I'm fishing the sea you know how I feel Riv Running Free you know how I
Free you know how I feel Blossom on the tree you know how I feel Blossom on the tree you know how I feel it's a new it's a new day it's a
new me and I'm feeling good dragon fly out in the sun you know what I mean don't you know
[Music] get get get get more do something about it because today we will learn how to and we did five on
Windows we launched a scam awareness campaign that prevents them from falling
the and when passionate people come together for a shared Vision the shall become the most loved learning platform on planet Earth and enable more
lives I think we are getting there here's my award and I'm so grateful to call because I wouldn't have gotten this award if not for the boot
understand hey I want to give a big thank you to theal and the code Basics team great work code Basics team I must say it is one of the best boot gam that is available on planet [Applause]
[Applause] [Music] [Applause] [Music]
another problem at [Music] [Music] made a problem that runs on a deeper
level and when passionate people come together for a shared Vision the world together for a shared Vision the world shall see a positive impact shall see a positive impact with you get
the basic because I wouldn't have gotten this award if not for the it's
five minutes so please have your [Music]
connect with volunteers I can see few people are s so can wall is help them out find them a seat
a powerful session we have a Ang and beautiful lady who is going to be presenting on technical things which is really hard for me but she is going to be presenting it is in a very easy manner so we have pry sing she is a
senior software engineer at agno an open source platform that empowers developers to build ship and monitor multimodal agentic systems capable of processing agentic systems capable of processing text images audio and video guys so I
request P Singh so stage is yours go ahead thank you you thank you for such a great introduction that made me sound way introduction that made me sound way cooler than I am yeah okay so okay I can
see the screen from here okay so hello everyone how is everyone doing I hope you're doing well and you're excited for the day the the way I am so a little
introduction about me I am pry and I'm a senior software engineer at agno I love to build things and I love to talk about them which is exactly what I'm doing right now and I'm super glad about this opportunity our today's topic is AI
agents and the future of work so this is one topic which probably comes up in everyone's conversation right whether you are Technical and non technical like AI will take away a job or you see some ads on Instagram from some chat GPT
experts you probably do right but what if I tell you it's not exactly what you think and AI agents are solving problems in real world we will be covering um all
these examples and use cases in my 20 minutes talk so so give me your 20 minutes and I'll convince you that it will not take your job but in still make will not take your job but in still make it more
technical term I wanted to cover a little story from my childhood right so there was this very particular Dosa shop in Bali Mumbai I grew up in Mumbai where my father and I used to go almost every single day but whenever we go we see
that the place is not filled with people it's almost empty and it almost always broke my heart because the guy made really good Dosa right but consider a really good Dosa right but consider a case what if the guy had an access to an
case what if the guy had an access to an AI agent where it can train where it can give access to all its processing all its execution data all its about its surrounding and find and have a datab backed information on when what and how
backed information on when what and how to uh how to sell products right now well this example is very very high level which may or may not come to life given the access but it for sure tells you that it does not replace the job of
the Dosa guy right so it's like having a companion how so I wanted to cover this one use case which probably all of us will relate to as well we are growing since
we're growing in a in a world of social media yeah consider um any social media manager in the house very wild question to ask um okay nobody that's great so I to ask um okay nobody that's great so I can assume a few things so I'm taking
this little liberty that consider a case where you are a social media manager and your boss comes up to you and tells you that hey our engagement sucks um sorry for using that word but uh our posts are not doing well we have very few likes
and almost no perment but your our competitors are doing so well how is that and he and the boss asks you to go and and probably get um probably create
a list of competitors and scroll through and get a strategy on why their why their posts are doing well and ours is not so this is the usual workflow that would that uh would look like right you create a list of competitors you scroll
through thousands of posts you take screenshots and you build spreadsheets and then come to the actual implementation and do and have a solution but but all the first three parts are boring who would like to do
that job like imagine going through thousands of post and you only spend little time on the actual problem but what if I can have an army of agent do all the work that I don't want to do like I I have I have it
want to do like I I have I have it created computors I I have it um I have agent for each of our website for Instagram for YouTube for Twitter probably and I have it create uh spreadsheets and I have it create
strategies for me as to why the competitors are doing well and you get to focus on the real part you got to focus you get to focus on the solution create strategies and do the brand refinement which is the important
part so this is a little snippet of code from agno and how easy it is to build from agno and how easy it is to build with agno so you you create a team uh you create of create a team of almost like an army of AI agents add is as
members like an Instagram agent YouTube agent and strategy event and you have a agent and strategy event and you have a team uh and team um leader overlooking all these things and do the work for you that's how easy it is to build with agno
just six seven lines of code but how does an AI agent exactly code but how does an AI agent exactly work and make all of this possible so I'll cover a little about an AI agent and how it works and then we'll delve
into a real life customer work that we at agno are solving so we have all these models right GPD 4 anthropic and they all are becoming so much better at reasoning writing accepting multimodels following
instruction every single day we see some advancement or some um some like instruction every single day coming up and which is what makes AI agent
possible but what if I ask the model right Char something like Char GP that well the Char GP application in itself is sort of like an AI agent but consider
I go and ask hey can you find me a free slot between my boss and me and book a call with us but the model gets sad it does not have any information about me does not have any information about me or my boss it cannot book a schedule G
but what if I build an agent and I give it some tools like you have access to my calendar and my boss calendar right now it can book a meeting for me because it has access to do but what if I make it
more powerful I give it a knowledge base what is a knowledge base is consider a PDF with my life diary suppose or something like any information about me
something like any information about me where it can uh uh take you take context from right and what if I give one more capability where it can remember our conversations remember about me that hey you are prey hey you work here your boss
is this you make that AI agent possible like this so you add memory knowledge and tools to the llms and make this AI agent possible Right you now you if you
give the prom hey find me a free slot it can actually output the uh output it for you and book a schedule for you right so this is how you make it and making something like that with agno is actually very easy which will cover in
um if you can meet me at the 7even floor and we can go through that but models are becoming very very smart every single day it just does not accept text
right cons like like you can see we can input text image audio video and it can also output text images audio and video it can see the images the way we do and
it can it opens up a plethora of opportunities and use cases for us this is also a little snippet of code us this is also a little snippet of code from agno um that's the case where uh
we've built an image agent which can understand and analyze images for you right you can have something like audio or videos as well in a similar and this is an image agent that can generate images for you now consider a case of a
social media manager where it needs a designer or consider a case where you have a photographer who wants to send you images where you you can put your face and you can get all the photos of yourself how useful that image agent
will be that's literally the lines of C how agno makes it possible so this is one my favorite definition of AI agent which is uh totally non-technical uh and it is AI agent
allow you to take the powers of the models and integrate them into your existing processes it's about Evol in our daily task and not replacing us so
yeah okay I'll get to something more Technical and something that we are supporting at agno with our customers I won't be obviously U mentioning the name won't be obviously U mentioning the name or using the exact data points but uh
this is something which is very interesting I love talking about it so we uh so we are supporting a customer where there is an admin it makes a system right system is sort of like a form which which suppliers are required
form which which suppliers are required to fill and send it back to the system to fill and send it back to the system um sounds simple right you make a form and the supplier fills it and suit very simple
except you're going to have many suppliers each supplier are going to have numerous DS numerous questions and you're going to be burdened with support right you're going to have so much manual tracking you you'll have to reply
to emails of questions you'll have to check forms if they they're right and so many things happening there's so much insuff inefficiency right but what we insuff inefficiency right but what we did at agno um my colleague aush is
sitting he's one of the um people who who made this possible by the way so uh toose to you for making it possible so suppose we have a supplier he has a
suppose we have a supplier he has a question he can reply to the email uh he can reply to the email where he's received the question the email agent checks the knowledge base which has answers to all the questions all the
contexts you would need it if it finds the answer from the knowledge base and the answer from the knowledge base and helps the supplier with the answer how cool is that right you don't need any manual tracking but that's not it what
if uh what if the questions are beyond the knowledge base what if the questions are very important that it needs to be answers right away the email agent is so
intelligent that it can identify if the questions are complex if the questions are something that needs to be answered it sends that question immediately to it sends that question immediately to the admin via email itself and this is
made possible through agno so how okay now you have a resolution the supplier can fill the form and you you have it but that's not it right now he goes to the questioner and he fils the
and he fils the form but you see the form is not form but you see the form is not complete or there is a little bit um the form is not complete the an are night right we have another agent which can
come I'm sorry can I just have some water
follow up all the questions and send it back to the supplier that's how um okay so that's how the stronger a system we made at agno all of this is made
the real use case I wanted to cover and we have got something exciting for you we have five really good um five goodies to give away but for that to happen you need to scan our QR um stars on GitHub
need to scan our QR um stars on GitHub you can make a post on uh link Lin or you can make a post on uh link Lin or Instagram but make sure to use these um Hashtags with it and I think one of our team members will contact five of you
for the giveaway these are very special goodies so yeah yeah okay on also there's insiders information this is not public yet uh we
information this is not public yet uh we have a $15,000 of haong going live next month so make sure you join our Discord uh for the announcement coming up please don't tell anyone I told you uh but yeah that's about it thank you or apparently
we have a few more Workshop tickets so I'm giving it away so make a post on LinkedIn or anything and uh we have five
LinkedIn or anything and uh we have five Workshop tickets for 300 p.m. so yeah that's about it and if you have any questions please meet me on the seventh questions please meet me on the seventh SW we have a agno booth yeah thank
fact about prti this is her first session you know addressing huge crowd isn't this appreciable guys at this very engaged that two hardest Concepts thank you fre it's really amazing okay so we are up with 15
minutes shall we play a little game funny and interesting game because we all would have got up at 5:00 a.m. I just wanted to know personally how many of you guys woke up at 5:00 a.m. or 6 a.m. rest all 8:00 a.m. then
came oh mainly wouldn't have slept I thought only Cod Basics did not sleep oh amazing okay anyways uh as we all belong to a data profession let's have a
to a data profession let's have a storytelling game right uh I data analyst apart from that I used to tell a lot of story at times I don't work uh I don't think uh my manager is here how kind of okay anyways at times not only
me many wouldn't have worked a lot at times we need some rest right but still you have to tell some story to the manager I've did that this that this right so we have to be very good in storytelling so we going to start a uh
small game with the storytelling let me start the first stanza it's have to be a kind of chain type everybody has to talk maybe we'll go with the first three rows done or random if you guys wanted to participate okay let me start like
participate okay let me start like this a very innocent girl from a village came to Hyderabad just to a in a career called Hyderabad just to a in a career called data analytics she went to Amir pit
tried to learn some to to fake courses thinking that that is the best course thinking that that is the best course ever ever okay done she has done with the course this is the story guys a village girl came uh sorry a girl came
from Village to Hyderabad then landed in Hy Amir p uh then could you please volunteers hand over the mics to uh that girl the second row this the
second row the first the story starts like this inocent girl from Village landed in Amir studed data analytics course next you can proceed anything she
met with an accident whatever good good any imagination started learning different courses oh wow interesting next during the process she understood that she's not much interested in data
but still she's interested in storytelling so she continued that oh facts then I guess maybe she's in code Basics event oh is that so yeah okay next continuation please finally she has identified all this as a hype and then
started and then started realizing the facts and the truths exactly then she facts and the truths exactly then she started her own coaching
YouTube and she now has 3 million subscribers now she's planning to start a Fest like this is happening here and she's
inviting people okay next good next good sir now she has a burn out and she uh decided to do organic farming my
can share her knowledge across and enable and so many people by opening her enable and so many people by opening her own channel oh is that so okay next she also started a business of um with organic farming selling organic
ventured into organic farming right wow okay wow am lot many ups and downs okay yeah make some tragedy guys it's going very positive note
tragedy some tragedy she wanted to expand her business but for that she expand her business but for that she wanted data data okay next she wrote to Donald Trump to come and invest in her
company okay now she is thinking of marrying someone as well because she this is what she needs now next yeah her husband suggested to hire some people and he gave her a very different idea she said
like whoever you are trying to recruit make sure that they work for free for the first six months free for the first six months sir in six months sir in strategy next she forgot why she came
for Hyderabad by by by looking at the likes subscriptions at the YouTube YouTube so she continued with that thank
you sir and then she realized that her reach is not going well with YouTube and she decided to make a movie on her life movie heroin and she ConEd she connected with rajam and they are in discussions postponing
my next and at the end of the day her mom asked her to woke up from the dream so it's all the it's all the dream wow so see guys how the story has
beg a girl from a village an innocent girl landed up an Amir pair then they've also married her then a business loss then some person has told me an interesting thing then movie okay this is amazing so interesting may know the
is amazing so interesting may know the time 20 ah we can continue the story movie sir next so I think she's in a Dream next so I think she's in a Dream Ride yes but you are know data analyst
you have to say some story continue whatever it may be okay she just woke up continue with the dream oh okay okay yeah uh she woke up and she slept again
again slip okay so she woke up realized that this is all dream so to start with she has to start from start strong fundamentals of from the beginning because it was a dream so she realized that in the dream
fantasy looks really nice but in the reality she has to learn and grow nothing comes free what an inspiration Koda s s but it's not only about data she'll be also having personal life right we do have we not
you know manager company data dashboard computer system friends no we do also have personal life something with personal life interesting think uh uh she didn't uh spend time going for trips so she went Bali after that Bali I
really wanted to go maybe after the marriage and there she was singing the song golden Sparrow enjoying in Bali golden Spar that's song okay super super sir next s
you're going to complete in 2 minutes please I know roll out she want now she you can sir
uh anything now she was in a dream right she woke up and came to the reality that I want to learn but she don't have a system so sad since she was coming from a village I agree true
a village I agree true unfortunately she get caught under D and him or whatever she know created all these things wow okay finally she has landed in a very good hands then on the day of her first training
some guy is following her on the last day of the training she is following next sir the other person she's falling a guy that is the end of the story
continue with the same sir why do you pass the pass the mic okay go ahead sir whatever it may be yes uh she already woke up from the dream and she's a UHA no no she wake up
already okay and now she's thinking that uh her one day is not enough for me so she's uh requesting God to make me a she's uh requesting God to make me a robo to do all those things Robo yeah
robo to do all those things Robo yeah okay tarun angal yeah so she realized it's better to be a PG owner than to get into data
Basics and she has opened her channel on to educate other people okay lessons
this much long long process I have taken it's better to manage a good why you take a long this process to settle my life itself I said go to the any matrimony site and get a good guy and just settle it out that's things
will be done very fast rather than all these things okay happy ending I'm these things okay happy ending I'm giving the happy ending got it got it so she's going to get married right okay next uh but right after that right after
it was another layer of dream she woke up
personal life right so maybe in that uh her personal life she just sleeping that's why all the dreams horrible okay okay so any final one good positive note maybe will close Okay three people okay you can sh you can go actually she uh
after marrying her she starts a YouTube channel that saying channel that saying hello guys welcome to my home
she realizes and writes a book of all whatever has been told and it sells as a compiles it and writes it in the form of a story and it becomes a bestseller ah
okay the next person someone a good positive ending yeah please yeah so after that book launches it fails very badly very poorly getting no reviews and uh then she goes on to Instagram and then starts promoting her Channel and
then starts promoting her Channel and then ask people to comment Ai and data then ask people to comment Ai and data fres for more hacks and oh okay so she turns out to be influencer in the end so that's how it ends thank you thank you
so much the you know the easiest job now is being an influencer on Instagram EG is being an influencer on Instagram EG phone EG filter a simple uh uh believe me guys people out there as an influencer they're making lcks a
influencer they're making lcks a promotion 15,000 15 uh 15 seconds G so anyways that is the best profession maybe she can choose because all this maybe she can choose because all this hardest Concepts why but anyways we
we have started somewhere and we have ended somewhere in a positive note anyways thank you all for this now okay clap for
now okay clap for yourself so comments okay uh we know the
11:25 AA okay what can't right but anyways already it's been there ha what else anyways we are
also going to be a technical session to tell about him he has traveled from Japan guys to address you all people this is amazing thing and uh I'll be this is amazing thing and uh I'll be introducing him in a while soon and uh
this day so what is in your mind so what is the context that you are having your mind you know for the whole day new learnings how many of you guys came here to learn something new
guys only came to meet new people very genuine okay how many of you people very genuine okay how many of you people came here to explore some new people came here to explore some new opportunities genu wow amazing and yeah
you're are going to take away at the end of the whole day and do you guys had any pictures with your friends okay did you post in link our
it's very uh easy to have all the descriptions done for your post and do tagers and the hashtag is AI data pH 2025 nobody in the crowd has posted nobody okay I'll give you the prompt and
I'll you I'll give you the description guys maybe post uh this session you can pause that do we have the speaker pleas back the stage okay uh so now we are going to
have a session on large scale data analysis in the age of AI and we have the speaker sa Saha and to introduce him he has 11 years of experience at NEC Corporation in the diverse fields of high performance Computing distributed
programming and data science he holds a bachelor's degree in computer science and is currently affiliated with the digital technology development Lab at digital technology development Lab at NEC as a senior research engineer so I
welcome sa Saha onto the stage and to add a point he has traveled from Japan to here just to address you people and to share the knowledge so I welcome sa Saha a huge Applause for him guys thank you so much thank you
you so much thank you sir the stage is yours
are excited uh yes so I'm excited I'm super excited and as he shared like we came from Japan so part of ncrn lab Japan and we are here to talk about some of the technologies that we are working and
some of the very interesting thing that industry is facing challenge with and probably something that will definitely help you if you running some business or doing a expert professional field of data science okay so without uh wasting
a little bit more time let's start with small introduction my name is SV I had my graduation from West Bengal University I'm from Kolkata basically and I'm having my career at NC for the last around 12 years by this time uh and
Computing data science big data and machine learning what is that thing that machine learning what is that thing that we'll be talking about today
motivation Tas form our main career that I have started my career at super computer that n had so my main skill is in high performance Computing at our lab I had my leader Mr zaka who is the primary author of the library that will
probably he's there in the booths that we have at 7th floor please do come let's interact so he thought of okay let's use our technology to speed up python what is the python that we can work on that is the time I believe we
started 2018 so data scientists were facing issue with the large scale data processing so we thought of okay let's build something that can speed up data science using the technology HPC technology compiler technology that we
have and that is when we started developing a library called fire Ducks probably some of you know about it I'm not sure because I had been speaking on it about in Pon India in the last pon pon Hyderabad also anyone from the
audience who already at least heard about the name fire du uh thank you so much of course okay so yes so I'll just dive deep into what fireex is and why probably might be able to explore that not learning but explore
because learning curve is very small it doesn't force you to learn a new library but it helps you to exceed the performance in your existing pandas okay so to start with that I have some of the quizzes and I have given to some of
these passes for the workshop probably that is a workshop that they're running uh after this session probably so I have five passes and I have many quizzes definitely but I can take five in order to uh pass this free one if you have
already not registered for sure but anyhow to start with that uh this is the regular workflow of a data scientist we all can relate although the trend is changed a little bit since the past few time but still it is same and we are
from several sources we preparing the data creating the model and it's not very straightforward we need to spend a lot of time during the data preparation creation of the features and all as for the research as
you can say as for the research says the most important part is data preparation and it stes around 75% of the ENT cycle starting from De loading preparation and visualization so a lot of research is been happening in this area in order to
speed up the data preparation before going to that we will try to understand probably we should follow then we'll explore some of the library in the field that automa the same then we'll explore something that probably you'd love to
something that probably you'd love to explore with okay so yes funders is one one of the dominating libraries so far and people are really enjoying it uh it's still very popular there are 100 million plus downloads over month but
when we go for large scale data it might not be a good choice for us the reason being first of all when pandas Source developed we don't have a concept of large data it is developed for someone who is very good for python but probably
people wanted to shift like help a python programmer to do data science not because that SQL is dominating for this market so python statist staticians wanted to use Python to speed up uh to do the analysis and pyth pandas was
is single threading so since it is single threading of course you can understand all they have very good syst it can become slow you have to spend a lot of time if you run something then if you're running on the cloud that is the
ideal problem of a corporative if we're running on the cloud we end up spending running on the cloud we end up spending a lot of cloud cost and we are attributing to the great factor to the environment because data server is
dioxide when it is running a long R execution so these are the ideal problem that probably we are facing the industry is facing and the research is going on although pandas has problems but the ideal problem come with pandas is the
way of writing when it is about the large scale data we need to understand some Bas practices and if that can be followed ideally we can speed up pandas as well okay that will be something that I will dive into today and I'll talk
about the optimizing strategy and some of the liability that offer the same okay squeeze time so I believe all of you know pandas time so I believe all of you know pandas right all of mostly at least 99% person
pandas Ro okay so this is a sample program that is trying to do simple thing it is loading at it has taking data frame as input trying to short the result by B column and getting top five from the a column both Fu and bar doing
the same thing let's now have a vote for according to you which method do you think is a good choice from memory consumption from CPU cycle from reability asset point of view can I have a vot for
a vot for four 1 2 3 4 5 6 7 8 9 10 11 12 okay great can I have a vote for bar 1 2 3 4 5
bar 1 2 3 4 5 6 7 8 9 10 6 7 8 9 10 11 11 12 same boat okay and rest of you 11 11 12 same boat okay and rest of you think probably maybe confused okay with
uh the V I believe uh uh one thing let's me let me have someone who has voted for bar okay sir okay so would like to tell me why do you think bar is a better approach okay first of all let me know
that is for the large skill data analysis have we yet raised no so probably I can offer you this pass then can you please tell me why do you think bar is a better approach whatever comes to your mind yes
filter is not happening here but you're very close may I know your name please sir sorry D danand okay let's have a big sorry D danand okay let's have a big round of applause for Mr danand who said
the key thing this is my data Frame data frame is a set of rows and columns ideally I may have many columns so let's see what happens in case of four okay
probably this is my data with 10 columns what I need I need to short this by B column and something like this will happen when I'll short it fine it's visible I think okay and what is a result that I'm interested I'm
interested only for the a column so when I say please dear p short it by B and give me a this will something be the result problem comes now during the the computation since it work on the entire data it already have
work on the entire data it already have spent time in shorting C to J that means materializing the result for c2j although we don't need to do that right so that is an waste of memory that is a waste of exception time what we can do
either as Mr danan said we can take a view of our data okay that means for me this is the two column of my interest we can reduce the data in the columnar
Direction and we can perform the same steps so it will now work only the portion of the data and I can save the execution time and memory consumption it's a very pretty common technique doing optimization like SQL Optimizer
and many other Library does the same it's called projection push down so ma'am uh let's maybe we can pass it on to Mr danand or maybe we can have round of applause for Mr Ganan for saying that and I hope you understand
the concept behind it and definitely I'll be exploring more and see how important it is when we deal really deal with large scale data analysis fine next with large scale data analysis fine next qu so here it's about doing a good way
of data flow look into the code what it's trying to do it is trying to get the male employee count of male employee from each country fine let's try to see
how it does two table employee and Country trying to merge it then and trying to filter those who are of male category then performing the group by and account so three male employees are from India and one is from Japan once
again please think here for a moment and try to find what might went wrong in the data flow okay because data flow is again important thing there is something
that is wrong from the performance point of view and if I may have yes sir uh yeah can you please uh so so so please uh yeah I just for also uh please can take the mic so here here we
merging both the datas then you're putting filter and then you're doing the aggregation so it's computationally intense so rather I would first filter on the gender then I'll do the merge then I'll do the aggregation great so
may I know your name sir Kosik Kosik so once again let a big round of applause from Mr Kosik have you register for the work sir no okay this is for you okay thank you so once again sir probably can you pass the mic too sir so I believe
you wanted to say the same thing yeah the same thing okay first we need to do the filter then we need to do the join we need to do the aggregation perfect so the problem as they correctly pointed it out is I only interested in the male
category why on Earth I do the data processing on the entire data I can get the filtration happen on the male category columns do the mods do the group by the simp alteration okay it will make your data analysis much faster
when doing on the large scale that is again something called predicate push again something called predicate push down now again I'm coming to you sir after doing this also there is something that can be done can you figure it
slide but there is something that still can be done to make it much faster little faster at least remove unwanted column least remove unwanted column because snit sir said SN sir right okay
says like what is the uned column for this data probably the ID because ID is something that is not needed and we need okay even if we remove ID okay we can save some p and time also anything else from the
data flow point of view look into the data flow this is an optimized version of the previous but still there is something that can be done and I have 15 or for 17 more minutes okay uh yes let's just think for a bit more seconds and
think if there is something that can be done can I have some bra and ra okay done can I have some bra and ra okay yes uh yeah yes the rate yes yes
the group by and then then we can merge it probably we can do the group buy and then March you will group buy based on what name name okay you will do the group buy based on name but what I need I need the
group buy to happen on the country right because I want how many employees are from India and how many employes you're very close but uh okay okay can you tell very close but uh okay okay can you tell me if you're very close I would say if
not then probably I I need to pass it to him okay yes yeah C code you have to group by on C code then get name of the country so but once again let a big ground app for your name please n my name is n sorry nisam nisam yes so he
said okay so I am doing Marge and then I'm doing Group by this is a data that can be country there are only two countries right so I can perform the
group by on this table based on the country code so for example do the filter then perform a group by okay for one there is three for two there is one now can do the mapping on the Marge one means India two means Japan or something
right and that will make it much more better because March cost will be less only two to 2 is to two margin will happen so that is once again the very important thing and this is the way I want you to I encourage you to
understand or think when we go for large scale data processing yes okay now let's try to explore some of the pandas alternative in this area and some of the
important one in this how many I believe spark is something that all of the present are here from the data science background spark you knows right spark we all know probably dusk how many of you know about
dusk wow great of course how many of you know about modin or at least have heard about the term modin okay Gan of course how many of you know about d d oh great and how many of you know
about polers poers skine of course he presented yes so now to understand that these are the pandas alternative and they try to solve the major two problem they try to solve the major two problem that is there in pandas one is not being
multi-threaded or multiprocess so we can use the available core or available nodes that is possible in all this Library second is query optimization by default pandas is eager what does that mean that means it
doesn't perform any intelligence whatever I request it to do it will do right after taking my request I ask dear pandas please load the data and short the B column it will load the entire data and short the entire data by it
doesn't perform any intelligence that means it's eager these are the library that does lazy optimization what's Le optimization I will talk in details in my workshop probably because this is very limited time but this the library
that solve this which one to use when the spark D modeling of the library processing if your data is too big to fit into the memory probably you can go for that spark is a choice for the SQL if you're very familiar with the SQL
probably you love probably you can think okay spark is for you and if you're from the pandas background dask and modin are somewhat compatible so you can find them good for you dug DB poers are the libraries which are doing pretty good in
the terms of single note computation again dug is something that comes with a familiar with comfortable with you can find it good PO is a python and it has compatible with pandas and if you want to speed up your
pandas code you need to learn polers and rewrite everything so a lot of migration cost is involved and that is the motivation where we develop fire du it should give you top class performance like in Duck Deb po although without
modifying any code it's that compatible with pandas so how compatible and all I'll just give a small demo because of limited time but in workshop I'll cover limited time but in workshop I'll cover with lots of hands on okay so okay okay
with lots of hands on okay so okay okay and I am left with with okay 15 14 minutes so let's have a quick introduction about fire it's uh the fire introduction about fire it's uh the fire terms as flexible IR engine IR is a
representation so it's a compiler for data frame we have compiler for language instead of executing a data frame operation it creates some kind of query and then it optimize it and execute it so this highperformance data
frame library with a compiler that does multi-threading that does just in time optimization it offer both the lazy and eager model by default it is lazy that means it can do optimization it can do multi- threading and it can do first a
execution the important thing is is of use if you know Panda's API you don't need to learn it it can easily plug to any library that P work with like say bur mat plot Li features first thing it's lazy okay so instead of executing
right after you ask dear fire du do this for me I will do it only when you need it so it will take a note of it when you need the result it will perform some optimization and the optimized code will be multi-threaded executed in multi thre
involment and that is something that will give you the lighten First Data analysis experience and and if you're running on the cloud the cloud cost will be reduced and if you are someone who is very passionate about the envirment and
like it will help you to reduce the carbon dioxide emission and if you love carbon dioxide emission and if you love pandas you don't need to learn new thing it's fully compatible okay why and how
again Workshop or you can come to my Bo I will discuss okay let's understand the usage so it's simply can be installed using bip instead of import pandas you can use fire du. pandas that is the only thing the rest your code will be same
now you can tell me S said me zero code modification so if I need to change the input how can it be zero for that we have some kind of monkey patching thing that is we call import hook so maybe you have a program with import pandas and
you have other module that is important and those module internally import pandas all pandas do you need to manually sense no just pass hyen option hyen name followed by Fus funders no code changes it can speed up your
existing program I will show you a quick demo for that for notebook or IPython Kel what's the way you can simply add this extension percent load extension fs. pandas your entire notebook can remain same and you will see the speed
up okay let's have a quick demo because uh I cannot have much time on this so what I did is I prepared something based on uh newwork set newwork taxi data said
and there are some sort of like you see uh I think it is visible let me still uh I think it is visible let me still make it a little Zoom okay visible a make it a little Zoom okay visible a little a little a little I think visible
from the back side also okay so it is trying to import pandas of course and it is trying to load a p file using the method called red P then it is trying to perform three query for example uh which is the most commonly committed vehicles
in the US state and what is a vehicle body type that is violation so some kind of pandas computation you can see it is doing the group by doing the shorting and chain of operation and on which day there are
happened and finally it is showing you the execution time so since it is written in pandas simply what I can do I can execute it something like okay let me execute and it is ex being executed using pandas
2.2.3 data loading is taking around 2.12 s 1 .8 query 2 1.86 query 3 is being executed and overall it will take around 11 or 12 seconds that's fine how can we accelerate pandas FAS pandas so first of all install it so I already have
installed so let's not uh so what I do I will add some hyphen name option right and this is hyphen name followed by fire du. pandas no code change just add this
and see what is happening so around 11.7 seconds become happening so around 11.7 seconds become 459 milliseconds so this is the speed that you can get without any code modification for fire from fire so this
is what uh a sample demo but when you talk about the Benchmark and I have 9 minutes okay okay so when you talk about some
Benchmark uh we'll show The Benchmark of course but let us also discuss okay so why Fus is faster faster than pandas faster than other alternative I'll talk about it as well if I get a little more time anyhow so first two thing we
first problem in pandas is solved using the multi-threading second problem of pandas is solved using the lazy optimization with the qu optimization third problem is it is written in very close to the metal in C++ with the
patented algorithm so all the algorithm the group by filter sh read everything is written in C++ from scratch and that makes difference how you can ensure that so for that probably this is a simple code this is a simple thing I am reading
a CSV file doing a short and projecting the column and getting the result to printed when this is executed in pandas it took around 46 seconds now what I it took around 46 seconds now what I said FAS is multi threaded FAS is lazy
can we make it single threaded can we make it eager yes how it can be make it single threaded by simply using the popular environment variable omom threads we can use and say one to make it single threaded how it can be eager
by default lazy how it can be eager just we can pass an environment fex Flags equals to Hy bench mode there are other ways also uh of course you can look into detail but this is the simple way now see this is single threaded and eager
see this is single threaded and eager this is single threaded and lazy single threaded sorry multi- threaded eager multi- threaded lazy this is the default configure anything this is something that you get and you can see the benefit
that you get and you can see the benefit so of course it is there but now can anyone of you tell me what happened probably what might happen here that probably what might happen here that make the eager to lazy to get such a
speed up multi threaded can be understood increase the core get it paral speed up is possible but what eager to Lazy brings that just sped up in this case anyone can guess okay for example consider I'm requesting this to
example consider I'm requesting this to you J sir ma'am please load this data sh the data by the amount column and find top 10 promises in column and save the result if you understand my entire query how you do okay will you follow one by
one instruction or would you like to do something that can be used to understand what is the lazy thing that happened over here anyone probably okay new face over here anyone probably okay new face of course yes sir
means okay remove the row that is not needed but are you talking about the row filtration or column filtration probably can you please be specific what can you please be specific what filtration you're talking about
happens over you but here is there any filtration in this code so I'm talking from this code perspective right hello short perspective right hello short of just a minute just a minute
also if it's not Clos then maybe you so yeah you have res one can I have M yes please uh so as that time is I'm thinking like there is a parall processing going on uh okay one minute ma'am do I have five minutes or okay
okay yes please yeah as the time decreased tremendously I'm thinking that the steps which are happening there are par processing so please see look at here so this column is for single threading this is single threaded this
is single threaded difference is this is eager this is lazy so no parallel processing here parall processing and eager is here only okay so this P up is from some optimization what might be that optimization Ser was
very close so I want if someone else yes yes prob that uh green sh yeah you want yes prob that uh green sh yeah you want you you the you you yes
before loading the entire data into the memory yeah of course why not because the data is in disk so I have even my freedom to do optimization starting from data loading so yes lazy means probably that right
you're not loading the entire data but just already starting to process it so so what would you do sir in that case you are again very closer to him but please I want the exact one exact answer and I have four minutes please okay no
can tell me now it can do the optimizing loading but loading of not whole data but part parts of data query that is what okay so first short values it has to rad these two column
the yes that was the answer that I was expecting okay since it only needs this two column which are the two columns amount and amount and SSN that is something that will be done here so for FAS and F
it will load entire data it will short entire data select SSN save it but when I'm lazy I have freedom to optimize I can see okay I only need amount and N so why I will load all the data I will load only those two columns perform the
shorting and store that is why it is much faster let's once again have big much faster let's once again have big round for nisam nisam I do already have shared right one so maybe this is for you sir who was very close I said yes
you sir who was very close I said yes may I know your name sorry Google okay so please so the last one and I have three more minutes okay do I have something okay there is something optimizing
feature a lot of features are there and with handson I don't have time to cover it here so just quickly let me tell you there are two layers one is the compiler another is the multi stting so compiler provide you the optimization like a
elimination date code elimination constant folding there is Dom specific data frame domain specific optimization there is pandas tuning parameter tuning or probably tried something that you would love to try anyhow this is
compiler layer in the back end it is multi- threaded of course it does manage memory very well using the of course and the Kels are retaining the patent algorithm so this makes it faster I can
take example and hands on off each of these in my workshop so looking forward to your presence there and if not Workshop come to my booth it is and 7 so inside I do have want to understand from your side as well okay these are some
benchmarks okay and these is a benchmark from tpcs Ben tpcs is very popular during transaction data analysis where we tried uh 22 queries using polar du fire du and how to read the graph this show the speed up from pandas so you can
see all these Library showed good speed up but on average Pol was 50 8 time faster than pandas duck DB was one and pandas duck DB was one and 199 time faster than pandas whereas fire
Ducks without any code modification can help you to speed up 141 faster from this Benchmark now what about the scalability I said that it is multi-threaded and I have some yellow card I think I just ick two minutes okay
okay since it is multi- thread probably you can ask me how does it scale pandas it's not multi threaded polar it scaled well but after some time I saw that there's a flat right and D FID scale very well with the number of code that I
system you can get very good performance of course from the multi threading and Benchmark again one of the popular Benchmark many Library Benchmark the algorithm on top of this thing when I compared FAS was on top of it we can
discuss more in booth and workshop with that note there is something let's let's keep at we you can come to both and because I don't have time I want to send something okay these are the resources website is there GitHub page is there we
can raise your query or whatever it is you can join us SL Andel would love to I believe all of you have the pamplet in your back that has all the links okay and we are running a blogathon campaign so probably if you are Keen about
writing blogs comparing the performance Library do participate here the top 100 something that is happening and I request you although there is not much time left but still you can do and you can share a beautiful Insight out of it
for today I have prepared something pandl so there is a simple problem that you can find by scanning this CER how simple pandas program is there it is not optimized I want you to think how you can optimize it the first top 10
uh this is open till 4:30 today so the top 10 good optimization from execution time memory will be awarded with some beautiful goodies with that note thank you so very much and I have one more pass probably someone who rais your hand
pass probably someone who rais your hand but was closed I said uh uh whom who who who who want to attend it can I have our hand okay the first hand come from there hand okay the first hand come from there so may I know your name
please okay so can you please pause this to him and that is the end I'm sure don't I have one let me take a keway and if there are a lot of question I can re them to come to my Bo yes do i is there any questions that you
want to ask me right now okay please Who who okay please sir uh so you are telling like five dos is very much efficient right are you using any uh gpus okay so there is GPU version as well okay but so far whatever benmark I
talk about is only for CPU because I'm comparing Panda so I don't want I want to make it fair so all this performance is based on C but when it comes to GPU the comparision target will be Nvidia QD can be very very less right when you use
can be very very less right when you use the hopper or black of Nvidia sorry when you use the architecture of blackw or hopper for this five dos it will be very very efficient right yes it can decrease the time yes so now it is better with
optimization and multi-threading on CPU with optimization on CPU it can run much faster okay so I would appreciate you come to my Bo and let's discuss it details anything else although I don't have time shall I make
it a close and ask them to come to my booth or you allow me to take one two more question one okay one more questions anyone yes
and API save why they are same so that is I talked right so I didn't when we thought of developing it uh we don't want the people to Bear the migration cost so you may have many large application that is
written in pandas that is what the industry have to be honest so when they want to migrate to any library that they need to modify the entire code so we mock the API down the line everything is different everything is C++ and all the
code is written differently there is no pandas dependency we just mock the API such that if you want to try fire du you don't need to learn a new API the same pandas program can be readily migrated and optimized that is the need we kept
the API same with pandas okay sir thank you thank you so with that let's have a big round of applause for all of you for the attendance session and do come to my booth and do take part in this and let's interact more thank you so
much thank you sir it is very insightful and interactive as well I can see many people were interacting with him and I I should agree to the fact that you did see you know drop a seed in many of our minds because even the concept that you
have explained is very new for me as well and uh thank you so much for that do you guys know what's the next session is we going to have a technical session right so we are going to have a
networking session again so take 10 minutes of time guys back to the uh seminar Hall in 10 minutes have a good time
complex topics simple and fun to understand for another problem at hand classes approximate 1.5
is job guarante something after he has been I still get uh email saying that you refer someone you get 20,000 we wanted to do something about it because we care about the humble
[Music] aspirants and we campaign that prevents them from falling for at Tech scams when I actually
didn't get a job that's when I came across sport Bas one of the coaching offering two legs I simply told him just don't two legs I simply told him just don't fall for the St just go for
fall for the St just go for [Music] the most loved learning platform on planet planet of I think we are getting
is all in our hands [Music] complex topics simple and fun to understand for
everyone but we realized that there was another problem at hand we en classes another problem at hand we en classes approximate [Music] a problem that runs on a deeper
a problem that runs on a deeper level job guarante been soon you get email saying that you refer someone you get 20,000 we wanted to do something about
it because we care about the humble [Music] [Music] aspirants and we did
[Music] w
[Music] w
[Music] w [Music] w
[Music] w
[Music] w
[Music] w
python 3.5 on [Music] [Music] Windows purpose drives an individual Windows purpose drives an individual Vision unites people
for a shared Vision the world shall see a positive impact with you the impact a positive impact with you the impact shall implify and enable more
grateful to call the basic because I wouldn't have gotten this award if not for the boot [Music] [Music] camp simply the best clear and easy to
[Music] understand hey I want to give a big thank you to Deval and the C Basics team great work Co Basics team I must say it is one of the best boot gamp that is available on
[Applause] [Music] planet hello today we will learn how to install python 3.5 on Windows [Music]
purpose drives an individual Vision unites
together for a shared Vision the world shall see a positive impact with you the shall see a positive impact with you the impact shall implify and enable more
grateful to call the basic because I wouldn't have gotten this award if not for the boat [Music] [Music] Camp Simply the Best clear and easy to
[Music] understand hey I want to give a big thank you to Deval and the co- Basics team great work Basics theme I must say it is one of the best boot camp that is available on
[Applause] [Music] Earth hello today we will learn how to install python 3.5 on
[Music] Windows purpose drives an individual Windows purpose drives an individual Vision unites people Vision unites people [Music]
for a shared Vision the world shall see a positive impact with you the impact a positive impact with you the impact shall implify and enable more
grateful to C the basic because wouldn't have gotten this award if not for the boat [Music] [Music] Camp Simply the Best clear and easy to
[Music] understand hey I want to give a big thank you to the V and the C Basics team great work Co Basics team I must say it is one of the best boot game that is
available on planet Earth [Music] python 3.5 on [Music]
[Music] Windows purpose drives an individual Windows purpose drives an individual Vision unites people
for a shared Vision the world shall see a positive impact with you the impact a positive impact with you the impact shall amplify and enable more
grateful to call the basic because I wouldn't have gotten this award if not for the boat [Music] Camp Simply the Best clear and easy to understand
[Music] hey I want to give a big thank you to hey I want to give a big thank you to Deval and the C Basics team great work Basics team I must say it is one of the best boot camp that is available on
[Applause] [Music] Earth hello today we will learn how to install python 3.5 on Windows
[Music] purpose drives an individual Vision purpose drives an individual Vision unites
together for a shared Vision the world shall see a positive impact with you the impact shall imp simplify and enable more
grateful to C the basic because I wouldn't have gotten this award if not for the boat [Music] Camp Simply the Best clear and easy to [Music]
understand hey I want to give a big thank you to Deval and the code Basics team great work C Basics team I'm say it is one of the best boot camp that is available on planet [Applause]
[Applause] [Music] Earth hello today we will learn how to install python 3.5 on [Music] Windows purpose drives an individual
Windows purpose drives an individual Vision unites people
for a shared Vision the world shall see a positive impact with you the impact a positive impact with you the impact shall implify and enable more
grateful to call the basic because I wouldn't have gotten this award but if not for the boot [Music] [Music] camp Simply the Best klear and easy to
[Music] understand hey I want to give a big thank you to Deval and the Cod Basics team great work C Basics team I must say it is one of the best boot camp that is available on planet Earth
available on planet Earth [Music] [Music] founder of code Basics thank you sir so next we have Mr nsh Shing founder
at campus X please come on to the stage sir thank sir thank you we have t sing is a mission learning engineer at Apple I request you on to the stage sir thank
the stage sir thank you next we have current Singh uh Kar you next we have current Singh uh Kar Singh Grover CEO Technologies thank you Singh Grover CEO Technologies thank you for coming sir next we have mega Aurora
founder and CEO at Dev Squad thank you ma'am for coming and lastly we have ma'am for coming and lastly we have sidhant P AI engineer at middle wave so let's welcome them all with a huge round of applause so we are going to have an
amazing session on data science thank you all stage is yeah all right folks let's sit down all right so as you know all this
is a panel discussion for AI and data science and you might have a lot of questions so what I did is couple of days back I posted on YouTube Community
I got some questions from people so we are going to use some of those questions and also we have some few other questions that we have prepared at the questions that we have prepared at the end you will also get a chance to ask
questions to these different people so we have set of experts here later on when you asking a question you can direct your question to a specific person and for that you will need uh to know about
that you will need uh to know about those experts okay so I will have uh these people introduce themselves so first we will start with mega so why don't you tell tell about yourself what do you do like what are your interest
do you do like what are your interest and so and I usually free uh frequent at uh
Tech meetups and conferences in Hyderabad so how many of you already know me or have seen me in events okay I see a few hands up so I'm happy so I Mega I have about 13 years of experience uh in engineering and then I moved to
developer advocacy and then finally I thought uh I'll just do deel as a service Consulting agency and uh as per Gro and chat GPT we are pioneering this
in India so uh do you know about developer can I make them yeah if you could just Define de developer relations for me
okay it's it's okay I I'll just answer it myself so I've uh learned to explain it in uh Layman's language so if you want to sell a t-shirt or a jeans right
you can do traditional ways of marketing but if you have to sell a software to a but if you have to sell a software to a developer you cannot say database so that is what developer ations is before developer marketing we do a lot of set
of activities to actually show the value proposition to Developers for any software so that is what developer relations Is that's
work as a machine learning engineer at Apple um my daily job is basically uh do research and development of new machine learning models uh my area of so I work completely in code research domain and my area of resarch search is uh
parameter efficient fine tuning I don't know if you have heard of it um also embedding models multimodel embeddings and the project that I'm currently working on is called what what we call a reversal C curse of llms so yeah that's
that's me uh hi guys my name is nitish and I run a YouTube channel by the name of Campus x uh where I put videos around data science machine learning deep learning and generative AI thank you
everyone uh my name is karandip I guess you some of you might have seen me in DB's post as well either on LinkedIn or YouTube uh so I'm the CEO atle Technologies and what I usually dabble with this businesses as well as the AI
team so people come to me and ask like I want to build this in AI or I want to build a product which is AI powered what I like to do is take them apart on that to it in my conversation which is coming up
next hi everyone I am sadhan I'm an AI engineer at middleware where I Empower observability with AI so my daily job is building agents that help users navigate
through their observability data apart from that I also an open source Enthusiast so I built small tools which are more mostly focused on learning so I buildt tools that uh any student or any Enthusiast can you know go through check
the code out and understand uh complex uh you know Technologies all right thanks for the introduction everyone we'll start with our questions now the first question is related to that reality layer that we
related to that reality layer that we saw in the keynote Karan deals with a lot of AI uh projects He talks with clients so I would like to ask you Karan
mention one or two projects uh that you are working right now in terms of the are working right now in terms of the technology used and in terms of outcome created I'll start with something which is very relevant to what a lot of people
are doing and I'll share how we are doing it differently so everybody has seen like how chatbot Works especially using LMS now the problem that uh him and D discussed earlier was there is a lot of issues with hallucination trust
facts and whatnot so we have a client which is working in uh communication space in us what their problem is they have built and shipped a chatbot to production now if you go and ask the chatbot how do I do this the chatbot is
simply going to send a link towards their uh FAQ probably and maybe not even like put in efforts to write it down like a human would now this creates a distance between the users and the business why the chatboard should have
like actually solved their query but what it is asking you to do is go put in some efforts would you like a chatbot like that no right uh so this is business recently what we are trying to do in this case is how can they check
do in this case is how can they check for every customer query how the facts or relevancies being answered by the llm or not and again like is is there a or not so these are a couple of things we are catering right now again the
implementation is around gcp uh we are using dialog flow apis we are uh also also using AWS to automate certain of those things again everybody loves Lang chain we dabbled a bit with it but we're shifting gears and using certain basic
open source uh I'm glad you mentioned dialog flow because I also get this question that llm has come and is dialog flow dead
llm has come and is dialog flow dead or J has come is statistical machine learning dead and the usual analogy I give is let's say you are driving in Hyderabad traffic or let me give you better example Bangalore traffic
better example Bangalore traffic okay so you're driving in a this crazy traffic will you take car or bike probably bik right it's lightweight it helps you navigate faster easy to maintain so there are
problems where you don't need to use gen you can use statistical ml you don't need to use swad to cut a lemon right you can do it with a knife all right now the same question I would like to uh have you answer uh in terms of like
projects you have been working on technology outcome created Etc um so in my previous company which used to be leveli it was working in the customer success domain so we used to get a lot of uh agent and customer data
so in that it is very tough to basically evaluate an agent right if if you are a company like uber who is like having uh thousands and thousands of Agents you just have like one two managers on top of it so evaluating every uh every agent
is very hectic and you cannot do that so what we were building is we were we we were not building we built it actually uh so we built a solution which given a set of rubrics which a company follows to evaluate an agent we were
automatically evaluating those agents for them so it was a very tough problem to crack in the first place because every client has a unique set of rubrics every question is unique and you're you cannot just keep retraining the model
billion model and try to retrain it every time so we'll have we have to basically train a model which was able to answer out of the box questions in the uh like directly on the go so it was trained on some data but is it was able
to answer on other data as well so I I found that very cool and uh we also filed a patent for it uh so yeah okay you want to mention any technology specific text you use so the technologies that we use again llms
because it's it was a question answering system so what I'm talking here is let's say you you you might have heard right this call will be recorded for quality and training purposes those those calls used to come to us and given that
conversations we used to uh analyze how the agent did so there were many questions did the agent greet the customer did the agent uh was the agent empathetic during the call did the customer ask for a supervisor escalation
customer ask for a supervisor escalation etc etc so the Technologies used was llm we used a mistal model uh it is an open source large language model and we' basically instruction tuned it uh on our set of data which we which again we used
chat GPT to generate that that data so that came in handy and then we Bas after instruction fine tuning we also did a RL reinforcement learning step where we aligned the model to follow instructions more using a technique called DPO so
yeah those are all right thank you for that let's move on to the second question and this got the highest vote in my YouTube Community post the in my YouTube Community post the question is companies are hiring only
experienced folks for AI roles this is the perception how can a fresher enter this field I will ask this question to sidhant because sidhant was a software Eng engineer then you you are an AI
Eng engineer then you you are an AI engineer now so what's your take on engineer now so what's your take on this so companies like do hire fresher but they need to see your credibility whether you have worked on some complex
projects or not so what you need is a proof of work so I would encourage uh freshers to participate in events and hack aons build projects uh and learn
how to scale them and then apply to companies so companies do reach out to you if you share your work and if you keep uh like build a build in public and build a public profile and keep sharing your work uh a lot of my friends that I
know from uh Twitter and Linkedin they have been approached by companies have been approached by companies because they write blogs they share uh whatever they have buil on uh like these platforms so first thing that they can
do is build a social profile and have a proof of work good portfolio uh push uh all your U your work on GitHub or hugging face and uh just showcase it to
the world I think you are referring to himansu probably when we were mentioning the blog Etc right he I think he's
just him uh that I know personally of yeah but himo is one of them yeah but himo is one of them so what's your take on that I I believe D and sidhant covered a lot of them but I'll start with a story so there was
this guy called fan who was uh working from Bengal and we hired him when he was in the second year now why he was building in public I saw one video that he posted on LinkedIn and he was uh talking about how he did this how he did
that he actually showed people like how you could do something similar this talks about a very big gap which is the gap of visibility if you're not visible you won't be hired we I was not even looking for that position at that point
but just because I stumbled upon somebody I reached out uh the team conducted the interview and we hired that person as sidhant said as D said the second Gap that you also have to overcome is mindset Gap a lot of
freshers a lot of Junior people even a lot of senior people don't focus on the mindset that they need need to have in like being an AI engineer I have had a they're frustrated just because their organization is forcing them work with
organization is forcing them work with Gen apis the mindset is to train a model Gen apis the mindset is to train a model to work on a tech and not solve a problem this mindset could also unlock a lot of opportunities for you so if you
have a right mindset if you are visible enough getting a job or say scoring a enough getting a job or say scoring a job from the market is not a hard thing and that's a natural survey into my next question for Mega so Mega you have been
an evangelist for community building for many years and you also work with some of the big tech companies Microsoft and Apple and mongod de Etc so for AI and data Fox I'm talking about aspirant as well as the professionals what tangible
community participation like attending this event or open source contribution K girl Etc sure so I think I'll talk about the mindset Gap and I had the exact words when you were answering and when you
when you were answering and when you said the word mindset oh my God uh So my said the word mindset oh my God uh So my answer has been stolen but no but uh no was just kidding sorry I'm a little uh sarcastic and funny and witty at times
uh so you have to bear with me uh but mindset Gap not only what uh gentleman mentioned here but also on the abundance mindset that is where uh India is the
fastest growing developer ecosystem but we are uh still you know not using open AI banking apis which my friends in UK are using we are treated as secondhand
developer audience still because I think our mindset has to go to an abundance mindset uh seeing the possibilities what is possible for example when I was
is possible for example when I was managing a Singaporean uh deel so some conferences or meetups they will be like are you coming directly from Singapore I was like no I stay uh in Hyderabad and I manage a Singaporean based out of
Hyderabad so they are like shocked they don't believe me they won't believe me and they would actually confirm uh from uh V who was uh my reporter and uh is
she your manager for sure and not only normal people but also my influencer friends so that is where the mindset has to grow a lot of things are possible uh to grow a lot of things are possible uh and a lot of things uh we cannot uh you
know know at home and there there are so many platforms but it's all scattered what is going on in Tech or what so that is where the power of community which I did not have in 2010 uh so how I am from tier 3
College how I got into Microsoft or apple when I always had my friends that companies I went to a lot of interviews flung those interviews flung a lot of Amazon interviews per se so uh but no shame in
failing because I could network with a lot of great developers who could tell me the resources that okay use this use that these are the things you need to communication background I joined Samsung as a fresher uh there was a lot
of IIT and nitens so that is where you know my mindset started opening up by the network so in 2010 I did not have these kind of conferences developer meetups in India happening you know happening in India and when I joined
uh as a developer Advocate I was so happy and I was like I was going to meetups and you know U I'm talking about uh 2 three years back uh very few females will turn up and uh that was my
passion that I have to see more females and even my managers from Dublin and then in us they were like uh get more females on the ground and then now I see the community I feel so happy and that's where uh that's where I started deil as
a agency as well that you know it meets my passion to see such so many people like I learned so many things as part of this panel itself like uh you were solving such Innovative people are
thinking about AI will take jobs but you can you were talking about what ex different fields AI will create what different roles so the nature of jobs will change humans will move towards more Innovative like you filed a patent
right so uh that is I think the communities will help open up that mindset and also help you leverage the network you build at these Comm with these communities and stay a breast of the trends that's it
yeah all right thank you for that so let's move on to the next question and this one I have this question I have for nii so Niti just in case if you don't know he has this YouTube channel campus X so one day I was searching for some
concept I stumbl upon his video I watched it and I was impressed so I have learned from you if you're not following his Channel
Hindi I love watching videos in Hindi so thanks for coming here the question for you is what are your thoughts in terms of projects people should be working on during the AI upscaling because that is
during the AI upscaling because that is the most common questions that we get would like to share something uh first of all I would like to thank you uh
Daval sir for inviting me here and I would like to share a very interesting thing uh for me coming to this event and seeing him uh physically is a fanboy moment uh I was also a student of Daval sir I used to watch his videos uh back
in 2017 I still remember you released a pandas playlist at the time and I learned pandas from Daval and back then you became my inspiration ke maybe I should also start an YouTube channel and I should teach like you so for me uh
seeing you like this and sitting with you it's a big pleasure and thank you for inviting me sir he's a very humble person he he has an art of transferring credit but he deserves everything that he has thank the credit for everything
he has bu okay now coming back to the question uh so the question is uh uh during your AI upscaling Journey how should you select your projects so projects are absolutely necessary you know that right uh I feel transitioning
into a AI job is a two-part job where one part is where you actually learn the skills you go and You Learn Python you go you learn SQL you go you learn ml algorithms and once you feel that I'm ready now comes the second part where
you build projects and show the world that you are ready to take up a AI job so select in a AI project I feel is very very critical so I will share a three-point framework how as as a fresher or as a beginner you can select
a project for yourself so the first pointer is ke uh I have seen this in many students that they have this urgency of you know completing uh the beginner whilea part and moving on to projects as soon as possible because
obviously there is a pressure to get a job uh I feel what you should do is when you plan a road map for your let's say you have planned that I'll study python for 1 month ml for 2 month and so on what you should do is you should also
Place small Milestone projects in between mostly people when they talk about projects it's mostly about a Capstone project big project that they will show in their portfolio but I feel smaller projects Milestone projects are
also important so what you should ideally do is when you complete a module let's say you are studying Python and you have completed it now don't skip skip over to the next part and start learning SQL or something else what you
should do is take up a small project let's say something like uh if I want to build a GUI that can uh communicate with a database a very small project but even by doing that small project whatever you have learned till now will solidify in
your brain the pointer number one is uh don't directly jump to the major project during your entire Journey spread smaller projects after every concept that you have covered that is point Point number one point number two is the
Capstone project the project that will get you a job visibility so I feel uh you should select a project that covers almost everything from your Learning Journey so one example could be uh let's say you are building a uh I don't know
something like that which will help social media creators uh understand their numbers their analytics so what you can do is you can build such a project where you will have to do
machine learning to do sentiment analysis uh where you have to do a lot of coding because you are after all building a Google Chrome Plug-In or maybe a website uh then you should not stop at just building the project you
should also try to deploy it which means you will have to get an understanding of how Cloud Technologies work how mlops work the basic idea is you have to work the basic idea is you have to select a project that should add all the
covered till now now the biggest question how do I select a project title so most of the questions are around this I'm pretty sure you guys also face
this issue can I get a raise of hand how many of you feel this issue what project should I build I I feel a lot of people will share this uh feel a lot of people will share this uh emotion so what happens is ke mostly
beginners they do very common kind of projects like movie recommender system sets and build projects now the problem is if I'm an interviewer and I'm interviewing 100 candidates there's a good chance 30 people will be
doing the same project so a good framework to select a project is think in terms of problems like research and find out ke what companies are facing what kind of problems are they facing like I'll give you an example let's say
s is a big company right and they are operating at a scale of India like huge daily basis now what you can do is you can go to their technical blogs and you can go and read what kind of problems are they facing for example back in 2021
I was reading their technical blogs and I saw a very interesting problem where they wrote semantic search which basically wrote semantic search which basically means user platform and he's searching
Biryani right but the spelling of Biryani is sometimes b y I sometimes it is b r y i and something it is completely rubbish compreh now you have just one entry of Biryani in your database if the user
types the correct spelling dishes will be showed otherwise there will be no be showed otherwise there will be no dishes recom now swi is losing business now the challenge is you have to make a machine learning model that can identify
machine learning model that can identify b y database b y database now this is a very interesting problem and sui is facing it now as a fresher if you can work on this and if you can build something sub substantial and you can
show it to the users interviewers the company they would definitely be interested to hire you so a very good framework to select uh projects is think framework to select uh projects is think in terms of problems key
blogs just search Uber technical blog and you will find out what they are you will get a lot of project ideas from there so I hope I answered the question yep thank you yeah that was a good answer so yeah t you want to add
answer so yeah t you want to add something so very very very great point at the end that uh for selecting the projects one thing I would just like to add on top of what uh sir already said is I think India is not really using
kaggle as it is meant to be um it has a lot of potential but I think I have seen especially in India people using it just as a data set platform it's not a data set platform it's a learning platform from where you can learn everything so
take me for an example I am just a mechanical engineer uh I go by the name of Mr Know Nothing which means I just don't know anything and I've learned every bit of every single concept from kagle itself you just have to use it
proper people are so afraid that I don't have gpus I that mindset problem that you mentioned I don't have gpus I will not win competitions bro you don't have learn you you know in in your lifetime you'll
not meet those people who are competing there they are phds from big big uh UCL UCLA they are from IV League colleges and they share notebooks all you have to do is Click simple for run their code line by line and understand what it is
doing so just I I would like to emphasize on this don't be afraid of kaggle no matter what the competition is go in at any point in time you have the
full power to analyze the data that will help you understand what pandas is what numai is at any time you can simply run a back propagation that will help you understand what pie torch is what optimizers is ETC so at any point of
time just take any any good notebook Fork it run it line by line try to understand what it is doing and make use of kagal and talk to people build a very nice community and it's the best thing apart from his channel for us and gav
apart from his channel for us and gav channels for us to really learn that's amazing Point tan and just to summarize so tanul was a mechanical engineer he was participating in Kel and he had some reputation like Kel He was a
great Kel Grandmaster just based on that he got a job as a ml engineer now he's he got a job as a ml engineer now he's working in Apple so Apple ml engineer having no formal data science degree or computer science degree so that shows
the power of online credibility give him big big hands all right now next question is for sidhant since you transition from software engineering to AI engineering I
would like to know what are the three tangible steps anyone can take to tangible steps anyone can take to transition yeah so a lot of software their job they know programming they know how to uh work with programming
know how to uh work with programming languages now at that time uh when they are trying to transition initially what should uh they do is they should just Tinker around with uh some models and Frameworks for example try open API uh
build a small project with it uh and like uh the speakers before me they have already uh mentioned how to learn and how to uh build projects uh Tinker around first and get the feel of it after you have gotten the feel of it and
bu build some small projects you'll be comfortable with it you uh like you can explore some topics as you go and as you learn and once like you have built three or four small projects now it's time to dive deep into it so uh revise your high
school mathematics a bit because like machine learning and a engineering is deeply dependent on linear algebra probability Etc so revise those Concepts and now you can uh dabble with pytorch and other uh libraries and learn about
the concepts thoroughly once you have these two what you have to do is you have to build a uh good large scale project that covers everything from data pre-processing to deployment once you have that you can like these are the
three steps now you can reach out to people or hiring managers who are looking for an AI engineer and share these projects with them that uh I have build all these things and that would be like make it a lot easier for you to
transition from software engineering to data size so as an AI engineer what I data size so as an AI engineer what I felt is uh AI engineering is only 20% uh felt is uh AI engineering is only 20% uh Ai and the 80% is all engineering so
when I build agents uh mostly what I have to manages the data flow how the uh I'm passing on the data to llms and how I am uh processing the data that is
generated by the llms for further processing or further uh iterations so that's my uh all right thank you for that on a
relevant note I see many people here today who wants to become AI engineer or who wants to grow in their career so there is something that Karan has to share on that topic so what it is doing is uh we are
relevant for people who have certain experience and who are trying to like Branch out into a consulting or freelancing career so what we are doing is you might go to spin wheels either on fifth floor or seventh floor you'll find
a uh QR code you could scan it and fill up the Google form if you're into Data analysis uh data engineering or AI you could like fill the form relevantly up and uh we would be hiring for certain
people in future in coming months but we also actively work with different consultants and Freelancers we also like from time to time get openings for interns as well so that would be your opportunity to a part of our talent pool
in future we also keep looking for engineers we have some internal tools and we are also making b2d platform uh so we are looking for some interns as well as full-time Engineers
interns as well as full-time Engineers so please uh get in touch yeah we have started hiring from hackathons because of that resume problem like most of the students are building the same thing so yeah I just wanted to cover that yeah
well thank you uh we also by the way hired so many of our code Basics uh team members through our resume project challenge so similar to Kegel we also conduct a challenge on our own platform and many of our team members who are
working with us they are the they are from those competition so winning those competitions can be a great way uh for hiring so in the interest of time I have still two more questions but what I would like to do is give you all an
would like to do is give you all an opportunity to ask the questions so once again I will summarize Oma is a community building evangelist she has community building evangelist she has worked with Microsoft Apple tanul is
with apple mechanical engineer to AI engineer transition lot of interest in research pile a patent nitish sir of so you can ask anything any questions
related to education side of it curent CEO of atck and sidan a Engineers so CEO of atck and sidan a Engineers so we'll have a open mind now so yeah let's open it up for the questions okay so we'll go all right so I would like to
take question from that person yes uh hello D sir I am a big fan of yes uh hello D sir I am a big fan of your teaching so uh is data structures and algorithm is uh required for data engineering data science
you are hiring a engineer at Apple you ask data structure questions yeah we do ask data structure questions yeah we do so it depends from um this is the advice I give to everyone and I think you'll agree on this that as a fresher you're
agree on this that as a fresher you're not really asked a lot right all we ask from people who are fresher is do you know what you know what whatever have you worked on be it uh anything is it unique and have you worked it on
yourself and do you know it completely end to end right it's not like we will ask you you know mlops you know deployment you know the model and everything no if you have just worked on a simple problem let's say Let's uh you
analyze it has to be unique to to basically let us know that you are good at it and if you have caught our attention all we have to look is is this project do you know everything from uh on it end to end for example uh at the
time when I was a fresher I build a thing called humor AI which uh basically tried to complete every sentence in in a joke form so any sentence you give it it it tried to complete it in a joke form so all I was asked about in all the
interviews was how did you build it what losses do you use etc etc there was no no other questions so if you are giving an interview for a startup or let's say a midsize company and you are specifically applying for a data science
role or a data engineer role or or a machine learning engineer role you won't be asked data structures and algorithm at at least that's what uh the current n is but if you're going for Fang or big companies then uh then the DSA is a must
so again that's uh that's uh how it is in the industry um so you'll have to like choose uh what you want to do but you can definitely get a machine learning job without DSA at least I got it okay thank you thank you you can take
next question so we'll go in the back the other person who is standing check the other person who is standing check sht hello okay you can ask the question okay so uh so my question is something like a for example so it's been almost
like a for example so it's been almost uh November 20 2022 right the charity came into existence but since then so we have not actually built a single llm model like India so still we are actually in the service kind of thing
and then see look at Deep seek and something actually is coming I Shar that feeling so that's a common feeling you and I thought that way and see there is there is a reason why that is happening mhm
mhm so we are still in a stage where we see if you talk about deeps right China China has advanced economically talent wise Etc I mean India has has a lot of talent but we
still we I think like we are kind of still behind like so if you think about China 10 years back we are at that place the research is happening there are like AI for bar right and there are a lot of other initiatives and we at code Basics
are also staring starting our research Wing so it will be code Basics research Wing so it will be code Basics research wing and we will be um uh you know building this initiative to research and collaborate with people across the indas
Academia private organization and individuals so I think that that feeling individuals so I think that that feeling that you have many people are having it and that will give rise to to almost this moment where we will see the next
model coming out of India it may take 5 years 10 years I don't know see the research requires a lot of patience and investment yeah so that I think that willingness is not there so if you think yeah apart from that like see for
yeah apart from that like see for example suppose when I search about any model like Mist trial and all kind of thing they do not actually Focus only on the the technical part like for example linguistic because the age journe is all
example semantic strcture what we're talking about there are many is one of the thing so which actually which is driving force to actually know understand the mission subjectivity so then that is completely locked here in
is only the technical thing but a is something actually which is apart from the Technics technical aspects like for example deeps when I search about them so there are many linguists so they understand you know the language in a
different way and they try to actually you know so that is hum for example expressions and all so this is completely locking in India like know this is what exactly I obser I'm observing it so I think you have some
take on it yeah I mean totally share your feeling but I hope that thing changes in next 5 to 10 years yeah sorry for the the same thing uh there's an initiative taken by
government of India recently they have started uh building the India's own llm they have started hiring and uh also I think uh a lot of for were around VC think uh a lot of for were around VC funding like how they qualify so uh even
that needs to change the VC ecosystem needs to change yeah all right I think we can take probably one more question I was just checking the schedule so I think we will end in like around 2 minutes and for
just in case I know a lot of many of you have question we are around here right so you know all the panelist they are here throughout the day so feel free to ask your question so we'll take one last question yeah like okay I think yeah
like few years back or two years back for a recommendation system like people used to use like mlr neural network and all so can we say in the future for recommendation system or anything people go to llms only instead of neural
networks or ml algorithms will we have the same amount of work which we were the same amount of work which we were doing in the I mean ML and deep learning like in the past like my question very good question so we Face the scenario
good question so we Face the scenario current right with uh with the project answer that question we faced exactly same situation we had to do text classification and we like okay we can just use llm and llm looks pretty good
in terms of marketing so let's say if you're a startup if you say I am llm jni powered startup you can ask for higher multiples when you are raising funds so we had a little bit of pressure of using llm but we tackle situation differently
Kar would you like to add something I won't probably add everything because my session is starting in 10 minutes and I'll be covering that but uh to like I'll be covering that but uh to like just sum it up llms are great but add
things they do you could solve certain text problems and they'll do it well but ask yourself is that good for scale that's that's the answer you should be seeking for once you have that answer you'll know recommendation
systems if you talk about it llm might be like just scratching the surface but if you want to actually make something for production it's better you go back to the basics and the other thing is the
interpretability with llm interpretability might be hard in some cases whereas if you have built simple linear regression model it's an equation easy to interpret in the industries like
healthcare and finance right we we discussed this briefly during the keynote interpretability is extremely important so he has a presentation in about 10 minutes he's going to go in detail uh he'll specifically talk about
that use case how we tackled it so if you have interest uh please attend this you have interest uh please attend this session all right so folks uh sorry we are out of time here we need to be getting ready for the other presentation
okay uh we are here like we'll be in fifth floor or seventh floor you will find us and we can discuss in person if you have any further questions thank you very much thank you all the panelists thank you
room the next session will start in 10 minutes guys and I think many of your questions as answered right right no okay if you do guys have
any questions you can reach out to them they will be here and the key takeaways that I got from the whole session is you know the mindset so even I was thinking from a different stream I'm from a bcom uh you know background so I was always
thinking whether it is really possible for me to get into AI so these kind of questions were running in my mind but I got the answer because I have seen a real example here who has made a transition from software engineer to AI
transition from software engineer to AI engineer and uh the other key items are how we can use kaggle and uh the other item that I wanted to share is uh we found two opportunities over here right they have openly stated that they are
actually looking for people internships so if you guys are looking out for any internship opportunities or full-time opportunities so you can reach out to them the primary one is ATL technologist the other one Mama stated so you can
reach out to them fill the forms okay we're all set uh I guess we have five more minutes left for the other session I guess let left for the other session I guess let me
check okay can someone tell me the time it's 11:4 so we are going to 1:4 we going to start at 1:10 okay cool uh may I know like who's the most beautiful person in the whole group beautiful wow handsome guy amazing huge
Applause for him he has ra hand wow is there any him he has ra hand wow is there any beautiful girl over here girl beautiful beautiful girl over here girl beautiful girl okay at least handsome guy no
nobody okay there is another guy wow amazing so other than those these two people nobody is the handsome or the most beautiful right you guys agree at that okay let me ask like this who is the most ugliest person in the whole
crowd you did not raise your hand for the beautiful person ra at least raise the beautiful person ra at least raise your hand ugliest person your hand ugliest person nobody okay so I just wanted to tell
these days we have been coming across so many incidents especially happening to software employees it employees so the mental strength uh mental strength is really very very important as a software Engineers we can see uh like you know
people are into lot many good roles but still their mental strength is not so so strong that I've been noticing all these days and personally what I wanted to say days and personally what I wanted to say in the sense in the job place we will be
facing lot many intuitions primary intuition is I have asked you a simple question who's the most beautiful person in the whole crowd and maximum like two or three people has raised their hands this shows that we are not like
this shows that we are not like accepting ourself is that true because we are been born in a society where everybody tells something AR if you would have put a little cream on your face M Muk straight little fit little
hair straightening all these things have been happening in our world right and that is where somewhere all those ideas got stick into our mind and we ourself works we may be very good in technical terminologies but not personally
mentally we are not so strong so I just personally wanted to say this accept yourself everyone in the whole crowd is the most beautiful person guys it doesn't depend on external appearance it is how you treat others or how you
present yourself your thought process your mindset are you guys agree with this if we are not strong enough how we are going to teach our kids it's it's not only like we have to teach them technical things we also teach them the
basic Factor they have to accept them whatever the hair okay no hair do no face do nothing but still you are the most beautiful person in the whole crowd except guys now
accept okay let me ask the question again who's the most beautiful person in the whole crowd wow so many hands days yes we are beautiful internally and externally
externally okay do we have the speaker ready okay now we are going to have a technical session
working as the CEO at ATL technologist with a passion for uncovering Innovative with a passion for uncovering Innovative use cases for AI and data currently brings a tech first mindset that enables him into seamlessly navigate across
domains so from Consulting startups on growth strategies to delivering Advanced solutions for smas is expertise drives impactful results so I welcome him once impactful results so I welcome him once again on the stage and yes sir the stage
again on the stage and yes sir the stage is yours thank
don't feel the energy by the way is it usually like this in Hyderabad like it usually like this in Hyderabad like it rained yesterday right it was quite for that let's do a simple exercise uh who who
a hand just a small
crowd all of them okay we're going to do something
up up down all set energy I up down all set energy I sub all right uh let's get started with the session now uh all right ignore the time please we are running quite late uh
I'll be talking about uh what is AI what ises a business care about there would be a case study as well and how do you been in been in now uh I'm the CEO at T Technologies uh
startups and some Enterprises oh sorry uh I'm the CEO at T Technologies and I advise startups for both Ai and Enterprises have worked with more than 50 startups so far and I'm also
Consulting some of them for like how they could like even shape their businesses by the way this is uh this is a QR code you'll also get at the end in case you want to connect with me on LinkedIn question what is AI
LinkedIn question what is AI anybody nobody knows what is a In This Crowd sorry okay anybody sorry okay anybody else insights of data anybody
else no one I'll go ahead AI is how predictive models or algorithms learn from data and produce outputs or input in a way that mimics how humans would have simply whatever you have told how many agree with this
AI for somebody if you are thinking technically it might be for a business technically it might be for a business for somebody like me it won't be a case for somebody like me it won't be a case now if we have to reframe what is AI
That's What I really like to say is it is just a tool when wheel was invented back when mankind was like just getting started with these tools it helped a started with these tools it helped a lot this was the same objective with AI
that is make life easier it is just a tool it's nothing else but just a tool so if you have to reframe the definition it is a tool that would help a business grow it is a tool that would help you create
it is a tool that would help you create better products and better startups it is a way to automate different tasks for businesses at the end of the day you may call AI anything it's about giving what
how humans think working like how humans thinks but to improve the business outcomes that is what AI is in reality all with me so far what is AI
all with me so far what is AI it's just a tool now what does a business really care about
a business is going to ask you all businesses care about his money of course some have a mission that is really Co to their heart but do you think a business could survive in this capitalist economy without money no it
won't even if you have tons of investment even if open a is having like how much investment do they have from Microsoft and other parties quite a lot right they still want it even if it is nonprofit they still want
a couple of things one being return over investment how do you optimize it you can increase the returns as simple as that you that you could reduce the investment as well
which is another thing you increase return you decrease investment simple math Roi is increased but in this economy you could also improve the market perception Market valuation around a company and that would be
beneficial as well now ai is a tool which would enable you to do want this session to be an interactive session by the way so if I ask a
humble request now we had uh client that we worked with they offered back office services for a lender like a bunch of lenders actually the client was getting thousands of requests they were
analyzing logs and messages and their task was to prepare some decisions out of it and like take some actions on it now they wanted to automate this now they wanted to automate this classification process simple
classification process simple problem now what would an AI engineer do else any specific models classification models
please OCR plus classification anybody else useing llm okay build a message classification model maybe using b or any other model maybe using b or any other classifier does business care about this
is data or not you need to build a classifier if you don't have data what are you going to do with it can you do anything without data in this anything without data in this world they have data tons of it actually
but do they have actionable data do they have labels around it with No Label you have labels around it with No Label you can't do
everybody has has read at least one machine learning article which is talking about how BT is the only solution to your text classification then your uh what is that emotional classification sentiment analyzer butd
is the only solution is what people talk about but it could be like you're bringing a rocket launcher to a gunfight you would blow yourself up as
this particular example with LMS like the gentleman was asking in the uh in the panel discussion so we started with an llm there were there were a lot of classes uh more than 100 classes in the beginning so we started with LMS we
wrote detailed prompts probably that is not a prompt that you would ask chat JB in real life you would simply go and ask like how do I classify this or maybe classify this in these categories that prompt was probably like 700 or 800
tokens long that is how we started by the way this this is just the the way this this is just the beginning now this was the final outcome we have a message we have a blackbox by the way and then there is
output does business care about what is in between no right do we so what do you think it could be inside this black box
inside this black box anybody segmentation again another guess yes back when the AI Evolution started
it didn't start in 2022 by the way AI is a lot older there are a lot of senior people that you could meet here like pan is one of those gentlemen over here and who have more than two years of experience in ai ai didn't start in 2022
we go a long way back even before this Century rule based systems still play a big role in simple problems even in complex problems why rule based systems enable you to reduce the inference cost for one second thing that they help you
with is making it much more explainable do you care about explainability of course everyone should second thing was the overkill Bird model
simply but this was handling 40% of data cases this was handling just 30% of it so you can clearly see we balance we treated of the benefits that rule-based
systems give and the inference that they like save you and we invested some of them in I guess there is some problem okay it's all so uh by the way we are in the world of fa and Technical problems are still killing us uh so we used bird
for some part of it next we used LMS if you look at this we are not building just single monolithic model over here we are building a couple of things together we are tying them up we started with LMS we then studied the
started with LMS we then studied the data we we then introduced uh like rule based systems we then introduced the bird based approach then we introduced llms it's not a simple task it's not like you go to a go to Google or chat D
and ask them you want to build a message classif classifier or a document classifier it's not how it works in industry in reality at least places where you actually create impact now there is another problem can
anybody guess what the problem was data model as well but how do we solve
that database is one thing we need label data we have a database but we need data we have a database but we need label need synthetic data but what if synthetic data is not as accurate you
have real data why do you need synthetic data you need label everybody agrees how data you need label everybody agrees how would you get to
relevant data inserted in right you could use Vector database you could use rag but unless you have the label data for fine tuning or any other activities you can't do it you have data the the problem is not data but having unlabeled
data I'll add on you could ask a client or a subject matter expert to verify this is it a good
idea why anybody wants to do this sit on a fine day review maybe 10,000 records and like do it again the next one especially if you are a business stakeholder would you do that if you're a subject matter
expert would you be going through this process how do we solve it please okay you have 100K messages you're going
correct the next thing would be you reduce it to 900 clusters this this is the real example by the way so I'm keeping the numbers realistic we reduced keeping the numbers realistic we reduced 100,000 messages to about 900 clusters
reduce it further but we wanted to make it an automated process we didn't want it to be something that can have some errors so first of all what we started on was we used a simple clustering algorithm we reduced it to 900 clusters
the classes were still 100 so there would be a lot of duplicacy but this is to make sure your client or subject matter expert has a easy life and you're not just constantly following the up on them for say three months you got to
make some money out of it right you got to make it to production faster now the human would review it and now you have golden records now you could go to an embedding model you could uh go to B you could do tons of things why now you have
golden records now you have golden data which is human verified working on a human verified data would actually make a lot sense you could have simply done something like this where uh say you have uh you have
built a rule based system or you have used llm out of llm you just consider the classes that have gotten out of it to be correct that would be a bad approach why do you expect LM to perform 100% well no right that is why you need
100% well no right that is why you need human annotators or human laborers in uh in place where they could help you create this thing which is golden create this thing which is golden records now what the impact could
records now what the impact could be quick WIS llm use Kia now it's easier take you a week if you're making a detailed one like we did but it would be
better than to wait for 3 months or 4 months for C client to label the data second thing could be turnaround time decreased this is a business outcome by the way initially you have humans sitting on the desk reviewing
logs adding it to different action items and getting ahead with it this system intended to save them some time the inference cost reduced how LMS are expensive right now if you're looking at it just from the API perspective it
looks inexpensive why because you're running it in sandbox environment the moment you take it to production you are making say a million requests you're going to see the real cost behind it by the way during the llm inference cycle
we are actually getting like uh bills of three or $4,000 a month even more than that in some cases but yeah efforts of course if you're saving client would love that this is a
business outcome that people are looking for revenues would increase why same for revenues would increase why same team more clients simple math more revenues the expenses of course decrease effort decreases time decreases you save
expenses the customer perception would also increase by a lot why AI power label everybody needs that label that you are an AI powered company but more
than that if you tell a client that you have sent me 100,000 messages or say a 200 ,000 messages it would take me a month to go through all of these versus you tell them give me a days time give me one hour give me couple of hours and
you'll get the result the customer perception shift in this case is huge and it would actually change the way a company functions and that is what happens now these are not the impacts that an AI engineer or a data scientist
usually looks at why we are technical we love to code we want to solve technical problems we don't want to solve business problems this is the mindset Gap that I was talking about the moment you switch your
perspective to have something like this like you want to save time you want to reduce inference cost you want to decrease efforts you want to focus on quick wins that is when you change things that is when you actually create
an impact and that is where your focus should be if you are in an AI team everybody can train a model everybody can use prompts to do things can use prompts to do things but can everyone create an impact in
business no right you have to change your mindset for that now let's talk about the final conclusion how do you win with conclusion how do you win with AI I could start with some memes but uh
everybody knows that you need quick wins you can't wait for two years before you get to a get to production how many of you have trained a model and it took you more than 3 months
people four okay some people are selective how many of you trained a model which took less than a week was it a toy data set no right but you know it yourself that it would be a rare occasion in some
cases when you have to productionize things you have to work a lot now in the interim is business just going to keep funding the AI initiative funding the AI initiative till a point yes after that no everybody
wants results focus on quick WIS I'm a CEO but I'm a techman so I won't ask this question but as a CEO of a non- tech company or even a non Tech CEO is going to ask for everything is this
AI the reasoning behind this is they could ask about it but it would be something complicated for them even if you want want to get your management stress you could work on the most complex model ever but still not get
buying how do you get it start with a simplest and knowledable approaches if at all possible again if at all because if you want to build something that is if you want to build something that is only possible with llms or say with
complex models you can't do it with simple rule based systems but if you can do it if you can make it more explainable do it if you can make it simpler do it every everybody was in software engineering at one point right
at least most of us there is one principle keep it simple stupid follow that even if you're working in AI focus on cost everybody of course like I told about a$ 3,000 or $4,000 bill for apis right does anybody like
that no because at scale it's going to be even bad what you have to do you have to make certain trade offs whoever is an AI engineer they know their salary already you are Prov family living in a tier one or a tier 2 City
the company is paying you a good amount if you're a Service Company you're charging a lot more to your clients but there is money at stake there are two types of trade-offs that you could go for number one being you give them a
quick win but inference cost could be high but it's still lesser than what you would be paying your AIT why this is needed less investment more returns second thing could be in the long term you can't work with this
approach right it's not scalable for the cost Elms are going to like take your Platinum they're going to run away with that as well you have to in in a longer term you have to work on to build a model which
you spend a lot more time on maybe a lot more money on to develop it but the inference is less that is how you work on production systems and create business impacts again uh I showed that uh that
in between usually when people start with between usually when people start with AI what they think about is use use bird for this classification problem why is that that is what sales that is what
sales in the market if you Google up right now how to do like how to solve a tell you that you may need to use multiple models nobody would tell you that it's not an easy job everybody's going to tell you use a CNN if you are
working on say images use a bird if you want to work on text use Simple approaches use uh uni model approaches basically that is not the reality basically that is not the reality monolithics don't work in many cases we
services client for another client we had to balance off three models one was another one was a configurational based model and on on top of that we had to
model and on on top of that we had to have some M like some static rules to reduce the false positives these trade-offs are necessary and hence single or monolithic approach might not work in many cases at least
might not work in many cases at least not in the can see a puppy as humans you could differentiate but if you don't focus on differentiate but if you don't focus on data quality and if you mislabel this as
a puppy and this as a cupcake you know what the result is going to be like right data quality Focus is something that a lot of companies lack the example that I gave with clustering I could have simply gone with a llm based answers
would have it been good maybe for the test maybe for the validation maybe till the time the AI engineer or the data scientist is testing that model out you would have seen the outputs what what about after
that maybe out this is a dog this is a sh cover but how many of you have educated your
your client any show of hands only one second a couple of them education doesn't mean tell them about the tech don't tell them tell them about how your bird classifier works or how your rules work tell them
what is possible with AI and what is not a client is always going to expect 100% it's not just like being a perfectionist but it's also about being a realist if an AI system
fails are there enough parameters or are there enough fail checks in place that life losses is there any compliance that you are violating if you are using a certain AI model if you don't know that you have
to know that why because you are also Tech Advocates inside your organization you you have to tell what is possible with AI you have to tell what is impossible with AI you have to tell where LMS would work and where they
where LMS would work and where they won't if you just keep on saying no or yes just because you don't want to educate your client you don't want to educate your stakeholders is going to come back uh come back and bite
come back uh come back and bite you that's the shortest way to you that's the shortest way to fail with that I'll end up with one code everybody loves technical Excellence but if it is not aligned with
your business objectives with your business outcomes it is just an expensive Hobby and and everybody agrees with me nobody is in alarmas sitting with me nobody is in alarmas sitting here he can afford we cannot Sam Alman
can afford we cannot so let's build AI responsibly with right things in mind thank you everyone you have have mind thank you everyone you have have been a wonderful audience
questions up in the networking session if you want yeah all right thank you everybody if anybody wants to connect with me here is my LinkedIn
session Splendid thank you thank you so much for that now we are going to have a delightful session any guesses guys uh which is the most a waited session even I was waiting from morning and I will I'll be letting you know a
few things during this networking session we have 1 hour 30 minutes with this 1 hour 30 minutes you can connect with your fellow attendees and have lunch and one more thing is we have our sponsors both on seventh floor please
use staircase and uh we also have spin games over there participate in that connect with them and as I've told there were like opportunities open at at Technologies and there is an other or right so please reach out to them have
right so please reach out to them have some great fun and uh don't regret later after uh wasting this time so you have 1 hour 30 minutes we'll see you after 1 hour 30 minutes at 300 p.m. in the same place thank you so much have
hello today bics is to make complex topics simple and
complex topics simple and fun to understand for everyone but we realized that there was another problem at hand we this classes
another problem at hand we this classes approximate 1.5 LS approximate 1.5 LS yes I paid
they said job guarante something after he something after he been get email saying that you refer someone you get 20,000 we wanted to do something about
it because we care about the humble [Music] [Music] aspirants and we did
that prevents them from falling for at Tech scams then I actually completed my masters and I came out and didn't hello today we will learn how to install today we will learn how to install python 3.5
[Applause] [Music] hello today we will learn how to install B
[Music] on purpose drives an individual Vision unites [Music]
people and when passionate people come together for a shared Vision the world together for a shared Vision the world shall see a positive impact with you the shall see a positive impact with you the impact shall amplify and enable more
lives here's my award and I'm so grateful to kod the basic because I wouldn't have gotten this award if not wouldn't have gotten this award if not for the boat
understand hey I want to give a big thank you to theal and the C Basics team great work C Basics team I must say it is is one of the best boot camp that is available on planet [Applause]
make complex topic simple and fun to understand for everyone problem at hand we classes approximate 1.5
[Music] L yes I paid 5 level job guarante something
[Music] after soon you get uh email saying that you refer someone you get 20,000 we wanted to do something about it because we care about the humble [Music]
[Music] aspirants and we campaign that prevents them from falling for at Tech scams when I actually
completed my masters and I came out and I didn't get a job that's when I came across basic one of the coaching offering him sports for 2 legs I simply told him just don't fall for the SC just go for
go for [Music] the most loved learning platform on planet
planet Earth I think we are getting
there is all in our [Music]
a [Music]
[Applause] [Music] [Applause] [Music]
hello today we will learn how to install python 3.5 on [Music] Windows purpose drives an individual Vision unites people [Music]
for a shared Vision the world shall see a positive impact with you the impact a positive impact with you the impact shall implify and enable more [Music] lives here's my award and I'm so
grateful to call the basic because I wouldn't have gotten this award if not wouldn't have gotten this award if not for the boat camp
thank you to Deval and the C Basics team great work C Basics team I must say it is one of the best boot camp that is available on planet Earth available on planet Earth [Applause]
[Music] down down
oh oh [Music]
oh [Music]
f no
[Music] f
m [Music]
[Music] down down [Music]
[Music] down f [Music]
[Music] oh
n [Music]
[Music] oh oh [Music]
[Music] a
[Music] oh [Music] down
[Music] down
[Music] down down [Music]
oh [Music]
f [Music]
oh [Music]
[Music] down
oh [Music]
[Music] oh oh [Music]
[Music] down down
f oh [Music]
oh [Music]
oh [Music]
[Music] down
oh [Music]
[Music] down he [Music]
f [Music]
[Music] no he no he [Music]
he down
f [Music]
home [Music]
[Music] a
[Music] down
[Music] oh down [Music]
[Music] oh oh
[Music] oh
down [Music]
[Music] now
f [Music]
[Music] oh [Music] down
n [Music]
[Music] down
oh oh [Music]
n [Music]
oh [Music]
a [Music]
f [Music]
down a [Music]
oh [Music]
python 3.5 on Windows [Music] purpose drives an individual Vision unites
people and when passionate people come together for a shared Vision the world together for a shared Vision the world shall see a positive impact with you the shall see a positive impact with you the impact shall amplify and enable more
lives here's my award and I'm so grateful to call the basic because I wouldn't have gotten this award if not for the boot
thank you to Deval and the co basic team great work basic team I must say it is one of the best boot camp that is available on planet available on planet [Applause]
complex topic simple and fun to understand for everyone but we realized that there was another problem at hand we this classes
[Music] L yes I paid they said job guarante something after completing the course he
something after completing the course he has been for soon you get uh email saying that you refer someone you get 20,000 we wanted to do something about it because we care about the humble [Music]
[Music] aspirants and we campaign that prevents them from falling for at Tech scams when I actually
I didn't get a job that's when I came across one of the coaching offering
performing on that on the basis of that we do have very interesting goodies as well are you excited everyone how's energy folks yeah how's the lunch how many stars out of
lunch how many stars out of five five three okay I know you some of you must be missing chicken and all the stuff I know we just thought you know happening right now with you know this all this virus and things happening and
uh okay how's the session so far all the sessions you enjoyed so okay let's let me ask a question how many of you networked like met strangers more than 10
10 oh more than 10 came to you oh charm oh more than 10 came to you oh charm BR okay more than five okay that's good
two that should not be your friend stranger H know okay anybody Network like 20 people I talked to 20 new people oh I oh I know okay that's good that's good so we
interactive quiz so for those who know code Basics we know that we do this quite often the slido quiz is our trademark thing to engage with people and uh so what you're going to do you're going to take out your phone and scan
the QR code right which I'm going to show now before doing that connect to Wi-Fi because I heard that some network is not having a good signal here probably who's using Gio Gio okay I think do you folks have
any issue with the network here yeah it's I don't know maybe it's some something uh so yeah please connect to the Wi-Fi so that you can play the the Wi-Fi so that you can play the quiz and
all right folks there you the the top five winners you're going to get a t-shirt good Basics t-shirt so there are people joining from online as well and uh you know you have competition from online as
two they are going to get a book a physical book from physical book from P it's a big thick book that's written on Azure open AI we also have the author here will be happy to sign the book and
here will be happy to sign the book and give it to you all right everyone scanned anyone facing a struggle
works if your connection Works let's give it a start just just give a just type something let's see if your connection
something let's see if your connection works I emu by check copy pasting I know copy pasting check copy pasting I know copy pasting is our favorite task to
the foood should be a catalyst it should you know kind of make people very fast show this to a cater saying like see somebody's feeling slow after the food somebody's feeling slow after the food after the lunch
stream is working well because I feel like the audio might not be clear please like the audio might not be clear please check once it's all
Indigo do you have an explanation for it why you wrote Indigo just like that okay Jil Jil jiga okay I see all the all the funny hyab
have been Distributing the badges from the spin wheels and stuff like that so how many badges you have collected so far let's see who is the winner six who got six anybody here who types oh you got six badges nice
zero okay 20 hey who who said 20 I want to okay 20 hey who who said 20 I want to check 20 bades yeah interesting I'll check later infinite
later infinite who
can't just lie okay in the next slide I'm not going to play the quiz is going to start actually so we have 115 people and uh yeah I'll just give one more minute for any of you to join in case you missing
from the next Slide the question number one so the way the quiz works is you have to answer fast at the same time you have to answer correct right the answering it first and who's who's answering it right okay you have to do
both fast's fingers first yes we copied it from this IPL
are in Hyderabad yeah I support Sr after
CSK yeah I who's supporting am I okay good luck see okay I think uh we have under3 people that's a good strength to start
okay the question number one on your screen oh no there is one intermediate just type your beautiful names let me see shank Arian neru jatin Dees Kevin snel bumika Danny ainash nain
Kevin snel bumika Danny ainash nain Shanker nice names gorov start on your screen don't look here look into your phone screen or whatever
screens you have question number one what does large refer to in the large language models what is the context of the word large
79% is it the right answer yeah you know popular answer does not mean it is right but let's check it's it's a trick question no I was kidding it's a right answer good job
was kidding it's a right answer good job who got it right oh okay 79% I get it okay all right let's see the lead of go yeah prik who's prik you nice
go yeah prik who's prik you nice satk oh you copied from Danny oh Danny Akash oh good
Shai Shai is not here maybe from online okay you one competitor from online Okay okay you one competitor from online Okay question number two
okay you have been hearing a lot of AI ml DL NLP hdf this and that what do you ml DL NLP hdf this and that what do you make out of
it so confusing huh oh that's a split po
the point is to get the most the best answer out of the four and the right answer out of the four and the right answer is let's take the leaderboard oh again the same fellas SAA
whoa good job Tanish oh I met you in the morning Danny is not done easier I guess okay all right question number three on your
screen so we need to have some questions on data as well isn't it
okay that's right yes it depends on the context and uh okay same FAS Krishna chananya ready who's that oh good n ISU this is the online F I guess okay
question number four oh there's AI B entered
that is the right option I'm giving you a clue oh I see divided opinions
here the right answer is parking good job good job parking good job good job all right SAA right to the top good job oh Shai Monica is here okay Sunil Kumar gurami
here okay Sunil Kumar gurami PRI okay good okay question number five easy one easy peasy you should all know this what is data governance
the data to convert it into gold can you please raise your hands I would like to see I would like to see your face [Music] online it's so easy convenient now to
say okay online online hey we can check the data you know please get the report I would like to see the faces of those people I was
expecting zero person there but it's okay you have decided to have fun now anybody in the 100th rank I would like to know who got the century today maybe later
okay question number oh s is still topping Ras enter the list question number six we are halfway through come on easy one
same two persons you guys have decided to have fun I know last bench folks I kind of suspect you no volunteers please check their no volunteers please check their phones online OKAY model context
protocol that's right answer good and the leaderboard Remains the Same okay now it's going to get intense question number seven number seven asset what does it stand
for on Google you can't use chat gbt in 20
seconds okay annual conference in tile same 2%
data I'm going to publish this on LinkedIn okay I I want to know the the two person Club really please raise your
five more to go question number AG what is Microsoft
fabric I know you're all waiting for the two person
in doubles you were 2 person before now you know gathered more fans and now it's 4 person I think two person Club is getting famous okay folks please check I really want to
okay folks please check I really want to see who they are hey subash I was joking you can't see there is no
Shai online oh online oh sad s is here Sunil Kumar okay sad s is here Sunil Kumar okay becauseas okay R is here okay good how
becauseas okay R is here okay good how to become a problem this question interesting no funny options
here okay 83% got it right and the leaderboard is same how boring question it is okay we'll go to the question it is okay we'll go to the question number 10 hopefully it's interesting
now 13% my God the trend is increasing like 13% my God the trend is increasing like crazy guys look at this that option itself is wrong but the spelling itself is wrong there is no word like
fitting yes if if you have got it right you know great good job and uh oh Ana you climbed to the top five good job she's uh I think if you're winning this you're winning it probably
for the third time if I'm not wrong all right last two questions the all right last two questions the semi-final on your screen
right let's see the leader board oh almost the same pkumar I think it's you okay that's great okay final question
final question the very last question on your screen
yeah I know what the two person is going to answer oh now 7% correct L chain okay to answer oh now 7% correct L chain okay 1% okay it's a easy one everyone got it right any guesses on who's going to be the winner
don't know who the person is and San okay otes for Sia all right is and San okay otes for Sia all right on the count of three two one drum rolls on the count of three two one drum rolls folks the winner is
Shai and Sunil Kumar and Pen great PR Krish Kumar and Pen great PR Krish is here okay another online okay how's how's the Jo now folks after this interactive
quiz you want more more quiz I'll do it later again great so now I'm going to leave the stage off to uh p and we have uh the you know the for the
first two winners n shivai and Sunil Kumar we have the author of the book on to the stage Cher it sh guys share for Aditya so Aditya is going to sign the book and give it to you in the code
please collect it from Adia outside the code Basics booth and the winners can the booth outside it'll be you like to word hi everyone
hi okay cool uh I'm ad saladi I'm a software engineer at Microsoft and I've recently written a book on Azure open AI um I think the top two ERS are getting this book today if you're not getting you can still buy it on Amazon so
there's an option for you and uh thank you so much for ending event and um uh on LinkedIn thank you thank you for coming thank you folks in about 10 minutes we are going to have a power packed performance from pen
really interesting he's going to talk about lot lot on llm and agents and uh just network with people in 10 minutes or you can just sit here whatever you prefer but be back by 3:30 3:30 is the timing right be back
oh oh [Music]
M Mr MK pan Kumar is a senior technical lead at openex specialization in Zen Ai and llm in with 15 years of experience across Erp Insurance Aerospace domains
across Erp Insurance Aerospace domains he holds dual master degree and MCA from osmania University and the master in data science from bits fan I welcome Mr ran Kumar to please be on the stage a round of applause
please I'll give you sign yeah sure you want to use this hello hi so hi so uh how are you enjoying the sessions for
morning great okay uh so what do you think of this is it a WIP coding no of course like me and my friend nothing it's not VI coding VI coding
nothing it's not VI coding VI coding it's love wife coding so um with that like uh I just wanted to tell uh I represent quadrant uh as an ambassador and then I've got I'm doing my research in uh uh lm's Mitigation Of
hallucinations uh apart from the introduction that I have got so without introduction that I have got so without wasting much delay let's go on to the agenda so first uh the introduction so what exactly uh we are going to do uh or
what you are going to see on uh uh the modern day software development because we all know that there are the wi coding tools like cursor wind Sur Klein and all
these things and how you code uh using these tools right so but that that's a different dimension of how you code but there is some other direction without uh disturbing the natural tendency or the natural flow that we are aligned from
last uh two decades of time in coding journey and uh the problem journey and uh the problem statement uh so uh like uh why actually
we have come uh um with these with one of the kind of solutions that we have been developing me and my friend and uh we'll be discussing that and then uh the or the tool that we are designing and then we'll have a live demo as
well okay so uh uh every H how many of you are working in how many of you developed chat bot agents or chatting agents customer
support or any kind of chatting agents good nice to see that so that means most of the organizations they have actually understood the requirement or the necessity of going into an agent uh because it has got some extra
achieve and now what happens in doing that is and now what happens in doing that is there is 75 to 80% of requirements that gets overlapped because we are ultimately developing a chatting agent
ultimately developing a chatting agent and agent and then you have a requirement overlap that's been happening so what are we doing here so we are developing the same kind of a wheel again and again
so do we need to do that no yes yes exactly so there is extensive duplication that's happening across the organization it's not internal to the organization that we are talking about and then you have no standardization
that's been put in place so if I ask if I choose randomly two person and then ask like how are you trying to develop your agents and then how are you trying to integrate both have got two different
ways of doing it nothing wrong but there is no standard of it right and and you have lack of agent reusability this is major problem right both are actually developing same concept but you don't have the reusability put in place such
that like one one of the major point which is covered in the previous sessions we get address which is that fast to market right because business cares about fast to Market in order to gain Roi so which is actually
lagging now the problem statement here we have identified in three different layers the problem itself self and then we have the
opportunity so in the problem statement we have rapid agent development and then overlapping requirements lack of Standards requirements lack of Standards duplication of
challenges then we have I mean since we have identified all these verticals of problems then to compensate or the Counterattack these problems we have Counterattack these problems we have something called centralized
integr process and then streamline development and then you have agent development and then you have agent reusability unified API access so once you have all these uh Solutions you can definitely address the problems that are
being identified and what what if you glue these two things then we get reduced Dev time which is atmost important for any business to make profits and then we have lower the cost and then you have improved quality you
have accelerated innovation and finally ecosystem growth so these are major important things for any business and for any software to make it to a peak level uh usability so how do you uh uh integrate
these three different layers so once you take the problem of Rapid agent development and overlapping requirements these two problems can be addressed by your centralized repository so if you have a centralized repo of Agents just
think of these are not normal software guys okay just think of a repository where you are actually trying to host your agent directly so some of you guys are coding agents so those agents are getting hosted in a repository as a
getting hosted in a repository as a central um uh what you can say is a a pack and then from which like other people can actually start consuming it without even coding and without even knowing it okay and because of which
what we can get so that centralized Depot will definitely and drastic Ally reduces your development time because you are not doing it you're just using it okay and then overlapping requirements will uh that problem will
get handled by your standardized uh integration and and your lack of Standards also will get uh uh addressed by your standardized integration and at the same time your integration challenges also gets addressed by your
standardized uh uh integration so this this is how we gel all these three layers so now you see this how we actually so and then your duplication efforts will get with streamline development and
then agent reusability and then your streamline development will be addressing your lower the cost and then improved quality and then reusability will be improving your improved quality and then
accelerated Innovation and finally with your API access you have accelerated your API access you have accelerated Innovation and ecosystem growth
okay so with that uh let me show the architecture how that actually looks like okay so we have uh a a two uh two module concept here which we call it as agent Hub platform
which we call it as agent Hub platform and you have something called uh traditional web application and uh in ag Hub platform we have multiple components which is which we call it as web interface
apis and then profile management observability and authorization so now these all components will make the agent platform
in such a way that it will host all the agents in a centralized Repository which will make the other party who is actually trying to consume those agents actually trying to consume those agents in a very seamless and scalable fashion
development of course in the over over the decades and then ages we we are habituated in developing traditional web applications uh that could be starting from your react uh angular nodejs in the
middleware or Java anything right and what if suppose you want to make that particular application a driven so it's not needed for everyone to learn the AI so you can just hook the platform pull the agents run it in your local and make
your normal application agent AI agent driven that's the major funa here and uh how many of you used uh hugging face Transformers come on okay so uh so uh
you know the Transformer Library which gives us the facility of uh downloading it right from pre-trained configurations and all am I right so very similar to that we have something called uh agent Hub client which will be you will be
putting that inside your client applications and uh there is there is a applications and uh there is there is a call that goes from your agent
and finally we get the responses back and you have two different personas here one is applic application developer and the agent developer so agent developers publishing the agents because they are pioneered in that particular area and
the a application developers will use the client library to pull the agent and to hook the application and then start utilizing the AI in their own application and own domain so that's how they get jelled up
and now uh coming to what what exactly is the uh advantages by using this particular platform uh and the approach is basically a four fold driven one is
apis which is seamless integration since we are giving a client and a server architecture you can seamlessly integrate and then pull whatever is required at that point in time and then you have centralized management as we
are speaking from the last two minutes and then you have robust and scalable and why is that because since everything is being now driven from the expert level definitely there is a concept of reuse uh robustness and scalability
because it's not V it is who is the expert subject matter expert that is the person who is developing it and now so we have something called interoperability this is most important and interesting part of this whole
and interesting part of this whole journey uh what if suppose um you have a requirement that you need to develop an agent uh but the features of that agent are not supported in one of the Frameworks of course you have number of
agentic framework these days you know it right just to name them you have agno right just to name them you have agno you have crew aai you have uh IBM B you have open a swarm you have llama index langra graph and all these things and
many more I don't know okay and uh so what's happening here so there is one features which is available in uh maybe agno and you're not seeing that in langra and something is there in langra you're not seeing in crew AI so what if
suppose I have two different uh agents which are very much popular and then uh developed in two different Frameworks example in cre a versus agno now you
can't integrate these two right because the entire semantics and the logic that has been uh developed is different so that is the major problem that we address here which is called interoperability so agent Hub never
cares about the plat uh the language or the framework of Agents it's uh framework AG IC so you can publish the agent in any of the framework and you can actually hook that hook the client to the platform and pull it and run it
and irrespective of the framework in which the agent is developed that's the beauty okay and now uh let's see the demo how actually it all works so we have three different uh layers inside that one uh it's very similar to your
hugging phase the concept is being inspired by the hugging phase where we can host the hugging phase mostly host the models and of course they entered into the agents but not in a full scale manner but we are doing it and then uh
uh you uh we see the UI where you can actually publish your agent and then you can classify the agent into different levels or a categories and you have uh uh apis supporting to that particular platform and you have a client which can
actually talk to these apis download it and then so it so we'll see all this in and then so it so we'll see all this in action
actually shoot them and then we can take that as well multitasking agent take
we take after this I'll continue and then we can take up okay so this is our then we can take up okay so this is our agentic platform uh so you see it right like we have uh a number of agents that got deployed and of course this is all
ize this maybe in the next couple of months that's what we are hosting and this has got two different flavors the Enterprise Gade and the uh the open source version of it I I had the open source Division and Shashank my friend
will host the Enterprise grade of it so we can deploy all the platform on Prem in an orc so and this will be uh streamlined across the different business units inside an organization and uh if you see it like I have uh uh
many categories and then we can search for an agent and agents are been classified versus within public versus private and then the private agents can be monetized if the contributor has hit a certain amount of threshold of
downloads and all those things so we are we are actually heading in a different we are actually heading in a different uh uh Dimension all together and
are running since you are seeing the platform so it's like the UI and then the API supporting them are both running in case and uh and and that's how the data is being fetched and then showed on onto the UI now if you see this uh if
onto the UI now if you see this uh if it's clearly visible uh then we have agent Hub as a as a predominant library in which like I'm actually pulling something called generic agents for financial task okay that's one of the uh
classes that we have and using which like I just uh uh I just mentioned which is called from preconfigured so generic agents for financial do task. preconfigured will pull the pre-configured agent but anyhow you will
be given uh the the flexibility to configure the agents as per your need you can change your the llm you can change different hyperparameters how many iterations an agent should go in in the background and all those things and
those all can be configured and finally you just pass in the agent name like come into your local very similar to an llm or a model that is coming into your local and then executing in in your local maion right and then let's see now
local maion right and then let's see now I just give uh yeah since Nvidia stock I just give uh yeah since Nvidia stock um analysis okay let's
see okay so now if you see that like I oh nice okay Wi-Fi
you said it's like centralized repository right so we have one problem with centralized repository in like in the software development like it should be distributed right uh so if centralized repositories missing the
connection because it is like in a centralized part right the centralized server so if we if can we do it is in a distributed way uh I'm I'm sorry can can I get the question one more time yeah I'm saying it's like in a centralized
way centralized repository right so we can in the software development we mostly prefer Version Control Systems like distributed way so that it it it will be not in a siloed way and we can be distributed the agent platform can be
in a distributed way in the cloud way it can be scalable also so uh that's good question just to repeat like uh your question is on uh can can can this pattern fit into a distributed nature instead of monolith or a single single
pattern right yes it can be done since it's just a repository and then a platform so once uh you deploy this code into a distributed platform that means your server would reside at one place where it gets downloaded and it can
scale horizontally as well so that's all it is it done okay
think it's a Wi-Fi problem so yeah if you see the just log just follow the logs I think it's getting connected it's downloaded the agent now it's connecting to open AI because I've just uh for demo purpose I've just configured it with the
open AI but as I said all these hyper parameters will be given for you uh in the method you can pass like from pre-configured you can pass all these tweak all these permutation combinations and it will work as per your requirement
and that's the whole analysis of the Nvidia you can actually cross thing so what I did I did nothing so we have a client and then I just hooked to
my platform pulled the agent and I asked it to run and then that's it so it'll run it for you guys and this is how you can have y number of agent what if suppose you you are developing an application and you need uh maybe you
are in a healthcare sector you need 1 2 3 four agents of some healthcare domain so it's not needed for you to code anything or you don't need to uh maybe anything or you don't need to uh maybe uh uh recruit the teams uh S I mean
which are having the uh expertise in that level since you are having this as an open source contribution you can download and then use it just similar to what you use uh your models as is with that like maybe uh how much time is left
for it I think like we are we are hitting it so we have 10 minutes um maybe open for questions now and then and then I we can take it yeah yes please I'll come to you so how do we Define agent accuracies like
it can give anything right like because it's a backend it's an it's a backend it's an llm yeah so um agent accuracy always depends on your llm of course uh there are multiple aspects that you can
measure with okay uh since we are giving since it's not tightly coupled or it's not bounded it's like open right as I said like you can always tune your hyperparameters in the methods that we are giving you so you can always uh uh
do this permutation combination of changing the hyperparameters and check uh the use cases against your own uh ground truth data set and then finally this is the parameter set in which like I'm getting either 70% or 80% that you
your production or next environments that you can promote uh yeah uh I just that you can promote uh yeah uh I just want to extend like what he asked for so like what are the parameters like how are we going to calculate like what
could be the score do we have any framework to test the agent accuracy okay so okay fine so uh to the knowledge that I have so we
don't have a specific framework yet to measure the agentic ACC accuracies uh but but still if you're it's all dependent on your use cases so if if at all you are using the agent concept for a use case of rag which we call it as
use the standard Frameworks that are like your DP vals ragas and all those things and then maybe Lama index also has got their llama index evals so you can use these Frameworks to do that and at the same time as again I repeat so we
have the uh configurations that you can pass into the preon configured method like what llm you need to use what embedding model maybe you want to use what is the iterations that an agent can go in so what is the top K parameters
that or top K chunks it has to retrieve so all these things you can configure by actually hook that to your own data set through data set and measure the accuracies and take a decision which
parameter set is best for your use case yeah thank you uh does this framework has a visual interface like langlow where we can design our own where we can design our own agents uh okay so first of all we are
not actually going by uh the no code uh platform we are not uh hitting that yesterday night itself we were talking maybe it's good uh to go there as well but it's in our uh road map but not right away uh the re uh the main
intention or uh the agenda for us is basically to streamline the process and make it more easy and has to get into the market that's our uh thing but as such as you said like uh uh checking that as a flow diagram or uh our
execution diagram so we don't have that yet uh the only thing we have is like then check like what are the agents published by you versus some other guys and then uh if there is already an existing agent which is suiting or
matching your requirement don't waste time pull it and then use it so just like the visual component we'll have a piece of code we have to understand what so uh example uh see let's say let let me take my example as an agent developer
and let's say my friend he's developing uh uh a use case for a healthcare industry so now he has a requirement of uh AI in his project now I just uh ask him not to develop that like just go to the platform and then check if there are
any agents available which is doing the same task and just put the client Library into his application and then pull it and use it on your server they run locally on your server yeah they are locally in your own
machines in your uh in your on Prem servers and fully controllable by you you can actually go tweak The Code by yourself because it's there uh how many of you know that your hugging phas uh models come to something called cach SL
hugging phase SL Hub inser very similar to that we have something called dot agent Hub of course I didn't show that but we the the platform creates that there so you can all it's all transparent you can see the code as well
so in your platform we can expect all the components that we see on regular L flow like what component uh exactly that's what I say right so we are not intended for you have to code the agent it like we are code first we are code
first approach they are like no code approach hello uh one second um sir would be here okay you can you can utilize that time into networking all right we'll be proceeding towards the next session if it's okay
sir thank you and then we'll be there here you can you can always come network with us you can know what exactly we are trying to do if at all you have a business case we are we are ready to help you guys thank you thank you so
help you guys thank you thank you so much thank you
found some people yawning is it because because of the because of the lunch did you eat
session post that will be having one refreshment so that you can again recharge yourself uh we have the session of the yourself uh we have the session of the data engineering playbook for AI success
SN uh snit uh Alam Raju is a director of data snit uh Alam Raju is a director of data and analytics at RSM us LLP with over 177 years of experience in the data analytics domain he specializes in data
architecture solution design and program management holding multiple certifications across data analytics Technologies I will I invite snail please be on the stage a round of applause
check can people at the back hear me okay perfect all right should we wait or should we get started we'll get started um because I'm
afraid if I wait more others will also leave my clicker uh yes
perfect yeah perfect thank you so much all right uh good afternoon everyone um I'm here with a topic that's probably a bit dry based upon what all of the other
topics have been here today but nevertheless I feel it's uh an important one at least from my perspective um before we get started you data engineering background and when I
say data engineering background DBS working with SQL Handling Systems getting their hand dirty okay good probably 20% of the people here how many of you are closely working with data engineering
working with data engineering teams okay what are the others doing in data engineering himself right so because you are preparing the data for
the models right so again you know we do have everyone today wants to talk about have everyone today wants to talk about models ai llm geni agentic ai ai especially right but what we kind of tend to forget is you know there is this
whole world underneath it which you should at least um I would not say identify but at least respect right because sometimes know you kind of tend to assume that the data is ready for the model but there's a lot of groundwork
model but there's a lot of groundwork that goes in um into preparing the data for AI model whether it is a language model whether it is a simple you know churn model or whatever that is right so what do if you are a CTO or if you
are a CIO or even if you are an analyst right there are certain aspects you know especially when you look at it from holistic perspective you know certain aspects of data that you know you need to be aware of and you know that's
probably what I'll touch upon a little bit today um and again you know hopefully uh it is useful and um you'll start respecting your engineering teams a little bit more because there's a lot that goes in or thought a lot that needs
you need to understand um rather than just saying hey the data is bad right so you are a decision maker these are some things that you need to keep in mind when you are thinking about even an AI strategy itself because
ultimately good models you know uh they about you know garbage in garbage out you know if your data is bad you your AI model itself will be bad and as we all know you know we talk about you know uh
compute intensive you know data sets you know where language models are being trained for days together Etc right we are all endusers of you know a language model or a vision model per se but you know if there are certain aspects that
are you know taken into consideration when you are training or when you're trying to set up AI model from a data engineering standpoint I think you know those are some things that you'll probably uh will make your life easier
probably uh will make your life easier you know rather than much easy right so how do you make your data AI ready and for that you know we typically look at the landscape here right you know if you look at uh the
traditional AI management I'll probably use the hand mic if you don't mind sir I'll use the hand
it from a traditional you know data management standpoint right you know you have your data cataloging data quality you know your data engineering scalability the data product everything that feeds the AI model but then again
then again you know once you have the data that is prepared for the AI model you know you have your metadata you know you have enrich data you know you have the semantics that are built in know you have synthetic data that is generated
we'll talk about it a little bit so ultimately you want to make sure that you are making your data ready for AI and you're making your AI ready for the data right because right now if you see there is a lot going around the closed
loop mechanism wherein you're are actually using AI to do some of your data engineering tasks as well right you know for example very simple example using llms to write your SQL code or someone talked earlier in the uh panel
discussion you know how do you use llm to write your Dax code for powerbi for instance right so you are using uh your AI to enrich enhance your data but before that step Z is to make sure that your AI your data is ready for the
your AI your data is ready for the AI right so when you look up uh let's look at uh three aspects of data engineering you know that we I would want to discuss today first thing of course is you know infrastructure right
of course you know the servers you know it lays the groundwork you know for you know scalable you know efficient AI systems by focusing on say architecture systems by focusing on say architecture storage compute Etc right so if you have
a robust infrastructure in place uh you know it helps in low latency you know High availability of your AI systems you know you'll not be writing a question for the chat bot and waiting for it to respond right so infrastructure becomes
really important you know when you talk about uh an AI model or a chatbot second thing of course you know the bread and butter you know the data right you know so when we talk about data uh there are three things that immediately come to
mind right you know one is of course the data quality second thing of course is you are building for the AI model to already consume right because of course you know a models can consume raw data but you
know if you're aggregating it you know preparing it for the model it performs much better you know has that has been traditionally been seen and then of course clean data you know quality of data you know in terms of accuracy in
of completeness right you know we have your typical data quality metrics uh it is crucial you know for you know accurate predictions or you know good model performance and then of course you know having again you know if you have a
know having again you know if you have a good a car you know you have good uh Parts but if you don't have everything that comes and puts it together your performance will be bad right so how do
you make sure that there is automation of your code there's continuous integration continuous deployment you know we talk about mlops Etc right right you know how do you take care of the process around making sure that these
data Engineering Systems are putting the data where the AI model is reaching out for right and know so that's the process aspect of it so let's go from uh
infrastructure standpoint first thing of course is you know how do you scale data course is you know how do you scale data processing right 74% of the companies you know based upon in a Gartner uh you know survey you know they have adopted
anyone of you worked with lake houses here data warehouse lake house data Lake for instance right so so there is a survey that talks about you know how do you ensure that you know you are using the right underlying uh storage or data
processing environment you know for preparing your data for the model right so someone again talked about I think if have anyone of you attended um Sor session in the morning in the fifth floor know he he did talk about how
pandas is there and and then how you use multi-cluster environment versus a single node environment right so if your data is large enough right of course you know distributed you know Computing is still the norm maybe we are moving away
from the world of you know Hardo and hdfs and moving more into the cloud if you know your regulations permit but again you know if you're not that big of a company you know today you know single node architectures are also working very
much fine you know they are working okay and they're giving you know high talked about dug DB for instance right you know they are it's a very proficient multicluster right so traditionally you know spark
distributed processing systems you know they were used for handling large data they were used for handling large data sets uh but you know advancements in know has changed this right so you don't need to um you know make sure that you
know you're using those large distributed systems and again you know elastic compute wherein you can scale up scale down as required you know it's a prominent feature of you know snowflake anyone work with snowflake here there
so you know how do you make sure that you're putting your comput sizing it at the right size based upon you know what the uh workload is I think you know the uh workload is I think you know that's always always there um so you
know making sure that you know you are deciding upon whether to use distributed computing and whether you want to use high performance in a single note because again keep in mind that any decision that you take any architectural
cost implication all of them have cost implications ultimately right but you know which one do you choose I think it is it becomes very important then of course you know how do you choose whether you want to use batch versus
stream right if you are training an llm model you know which has large volume of data typically you know batch would work best but you know if you are working with a chatbot model right you know where you need to have
interactivity between the end user and the model and wherein your data is more streaming in nature right you know that's where uh which is very real near so that's where you know streaming data comes into picture right so batch
processing is typically ideal for you know large data sets historical data sets which you train the model on and streaming is primarily good for say chat Bots or you know detection per se then of course you know standard you know
classic you know database related you know how do you make sure that you are putting in your indexes you know you're putting in uh your partitioning of data to make sure that you have increased query performance and again these are
things that you if you are an AI engineer or if you are a data analyst opaque to you because you know you are already having that data massaged and prepared for you but if if you are someone who is really interested in
knowing the underlying aspects of it which I think every analyst should right you know similar to how a data engineer should understand what the analyst does or what the person on the front end does so that they can incorporate more best
their systems I think the other way know you as data analyst or AI analyst need to understand or even AI end users need to understand these aspects of data engineering all right uh so you know how
partitioning you know how do you make sure that you're indexing for your efficiency uh for Effective you know retrieval of data right and it it actually impacts your model performance and how fast it can retrieve the data
right um so ensure that you're planning for scale right you know you may start small but you need to ensure that you know you you know where today you might be having 100 users tomorrow you might be having
10,000 users today you might be having gigabytes tomorrow you might be having petabytes right so start small start based upon what your use case is because of course there's a cost implication but then again plan for scale right you know
how do you make sure that you are able to expand it uh based upon your workloads then of course we talked about compute but then of course there's also storage right so again we talked about you know how do you make sure that you
are using the right technology for The Right Use case right first of all simplify you know if you are using just bi specific systems a data warehouse would be more than sufficient right you don't need to have
top of that of course you know most companies do that today because you know it's cheaper of storage costs I think are the cheapest among all of them right you know storage literally comes at cents so
it's very tempting to throw in a uh data Lake out there and push all of the data whether you need them or not and push it out there if you're able to use it make good use of it fine you know as long as you have guard rails in place uh to make
data you have some archival mechanism in place you are putting in together uh regulatory aspects of it that you know you cannot store certain data uh in certain regions and all of that hype on gdpr Etc regulations right so make sure
storage based upon your workload you know if your workloads are just about uh having Downstream implications for the data whether it is just sending data Downstream or just utilizing it for reporting I think just using a data
reporting I think just using a data warehouse on a cloudbased uh database or on top of uh lake house also it's fine but also depends on the nature of the data right you know typically data warehouse kind of uses structured or
semi-structured data right but your AI workloads would require a lake house because most of their use cases are around semi structured and structured data which you know you might not again there are use cases where today there
are workloads that can also use uh on top of the data like do querying for the unstructured data as well but not a typical use case right so uh make sure that you know you are having these
individual simplifications But ultimately as an organization right you you might also want to look at a unified approach you know sometimes you know you might want to have uh a single place where you know
there is uh consistency in terms of data because you also need interoperability you know there might be um data that you get from an Erp system which is part of your you know data warehouse you might be getting streaming data which might be
part of your lake house now how do you want to make sure that you are combining analytics environment right or for your for your video analytics model so you need to have a unified approach you know when you want to do that then of course
you know file formats right you know um these are you know typical file formats interoperability irrespective of the cloud platform so all Cloud platforms today you know they understand know par AO and or uh orc even Json you know for
that matter right so it enables you know seamless data exchange you know between systems you know it ensures that there's compatibility uh across you know diverse platform and makes it very you know Cloud agnostic right and again you know
park for example it's columnar in nature right so it speeds up you know your processing uh and make sure that you know data is much available faster even though it is at scale right and then of course lastly there is versioning right
so there might be use cases wherein you might have to go back to a past you know for an audit requirement or you know you might want to um query data from a previous point of time uh whether it is for debugging whether it is for audits
right that's where you know your time travel you know makes it easier because AI models might you might want to have that time travel built in into your AI model and for which you know you need to have time travel queries and then of
right you know so you make you want to make sure that uh you are using Technologies like you know Iceberg or Delta table which can enable uh versioning and which can make sure that you know you're going back to the past
and being able to retrieve data at a specific point of time right so we talked about compute we talked about storage now let's talk talk about you know some architectural considerations now now again this is
beyond you know once you have your storage and compute in place you know you want to make sure that you are once you are having your model or AI model connect to whatever underlying infrastructure or data engineering
architecture you have you might want to modularize it right you know what are the advantages of you know having microservices for each of your workloads and any thoughts what are the advantages of having
individual modules uh for each of your individual AI workloads any thoughts yeah sorry save some time lesser time yeah faster right you know because there is no dependency on each other and each of
them can be best-in class and need not be from the same platform right you know you can have you know a quering engine from one platform versus a reporting engine from another platform and you can combine them together like like like a
Tetris right so then of course uh you know you want to make sure that you are having that modularization in place because you can easily plug andplay you know aspects or features as in when you require you know Implement your you know
API you know observability tools you know you have um this way you are making sure that you are if there are errors in your model or if there are bias in your model for instance you know which is fairness in AI models is a very big
might not even be able to detect you know without having having an know without having having an absorbability tool in place right um so you know you have you know platforms like you know data dog you know New
Relic um there's also another one uh Prometheus I think right so these are different absorbability tools you know for your API you know making sure that you are able to log uh detect and you know correct you know issues in terms of
know correct you know issues in terms of your uh errors from your apis right and response times ultimately right and then of course you know making sure that you are already considering apis you know for you know that support both batch and
individual ones of course you know depends upon the use case but typically uh apis that support both batch and stream you know they help in handling both large data sets uh and also helping you know real-time data flows so for
instance you know you have you know Apache Kafka uh you have your Azure event hub for instance um you also have AWS skinesis or you know or data breaks data Lake all of these are examples that can handle both um real time as well as
you know your large data mod so how do you ensure that you know you're keeping it uh one size fit all but making sure that you know you're using best in class that you know you're using best in class right so um lastly ensure that know you
have endpoints integration across multicloud you know what this helps you to do is you know if you use say cloud agnostic platforms like you know how many of you have heard of terraform right so you know using the
these are you know Cloud agnostic you know you can use terraform to deploy to any kind of you know Cloud for instance right or you can use unified you know API getaways you know or getways for instance you know for so these actually
help in interoperability you know across multiple uh Cloud systems and again most of companies today big or small right you know we are logged into at least two Cloud platforms right you know you can use snowflake plus Azure gcp or AWS plus
Azure sometimes if you are rich enough or if your data teams are not good enough you know you will use all of them right um so ultimately you want to make sure that you are having something that is cloud agnostic and helps you to make
sure that you are preparing your platforms for deployment for better you know AI modeling then of course classic these are some things that you know you would have seen in machine learning or in data
know AI as well you know how do you make sure that you have well defined you know features right you know they help improve you know model performance they reduce the complexity you make sure that know you handle no real world no data
know you handle no real world no data issues much better right um so when you talk about uh feature engineering you know some things that come to mind is um transform the data how do you aggregate the data which is at a much granular
level so that it can be fed uh to the model right again most of the times you know features matter more than the model itself than the algorithm itself right so having a feature whether it is part of the data or whether it is engineered
you know for the model itself I think it makes a lot of importance of you know how well your model performs and how accurate your model is right so it impacts so features impact the quality and performance of the a model so again
typical um Transformations right you know how how many of you understand normalization so you know you scale 0 to 100 you know so you make sure that you are putting everything at at a normal scale know categor categorical encoding
know Group by Rolling windows again typical you know SQL based um you know c um features you know that you would typically do right and then of course I think this is something that is uh always uh coming through especially when
data is less right how do you generate data that you don't have and which you feel that the model is um looking for right you know sometimes you might not have a use case or you might not have data for a use
case and you want to make sure that you are generating that data again there's a lot of you know hype around this uh synthetic data generation and you know synthetic data generation and you know some there are you know two uh class of
data is useful one who say synthetic data is not but and I'm of the opinion that if you have no data you have no choice you know you do have to gener that synthetic data right so Gartner projects you know that by 2028 you know
80% of the data used by AIS will be synthetic right up from 20% that you have today right so what is synthetic data you know it's also called bionic data you know in certain areas you know it is typically generated by AI based
upon inputs that you give from your actual data itself again there might be you know multiple use cases on how this data is generated you know some some data is generated you know some some models tend to look at uh the existing
data and extrapolate it based upon you know how the population looks like um again you can choose what features you need to extrapolate that and generate multiple ways to generate this right so
bionic data you know essentially is how do you add new data or AI generated data to real world data so that you know you can train your model for scenarios larger scenarios right um so you make it more realistic you know by adding AI
generated data um especially in scenarios where the actual data is limited right and then of course hitl you know human in the loop you need to augment you cannot always I think we have been speaking about this from the
have llms they can generate the data once you feed it but they don't have that intelligence yet right to think whether it is the right data or not so
that's where you as a data analyst you as the subject matter expert know you come into the picture and to understand whether that data that has been generated that synthetic data that has been generated makes sense from a domain
standpoint makes sense from a model standpoint right you know uh if you have age as a column and if you have a population that is less than 25 80% of your data and you use a AI model to generate new data which has age from 25
generate new data which has age from 25 and if it generates a age of 125 right maybe makes sense maybe not right you know kind of an outlier you know I you know kind of an outlier you know I would I wish I would live 125 years or
we wish we had more data that has 125 years of age but so that's where you come into picture you know is it really correct right so you want to make sure that you are putting the human in the loop to ensure that you are uh
interpreting that synthetic data T has been generated and you know also of course making sure that you are uh correcting it right
metadata you know data lineage right you know how do you make sure that uh you know how do you make sure that uh you have data lineage from an understanding standpoint and where do you guys think you know this lineage is very
you know this lineage is very useful governance absolutely right so tomorrow you know you might have an end user that says hey your AI user that says hey your AI model gave me uh a comment that I don't
find appropriate right I'm going to sue you right now if you as a company who owns that model needs to understand why did it even generate that you need to have a tracking mechanism to understand okay
what where did that actual data come from why did it actually give you that output how did that inference work right you cannot have it as a blackbox right so that's where making sure that you have that metadata management inbuilt
actually feed the AI model you know becomes important right so if you are having metadata tags you know you are using for indexing and searching one helps in model performance you know if you are adding Rich metadata to your
models right you know it helps in model tuning one of course it helps in better feature selection because you know if you have metadata around those features you know it helps you to it helps the model to select the right features that
are required for that specific performance right then of course AI model rely heavily on lineage right you know if it so as I as I said you know origin is important because it helps you to understand where the data is coming
from and why the transformation has been applied You Know It ultimately it's about as someone mentioned governance you know you it helps build trust with the stakeholders the stakeholders can be your AI users the stakeholders can be
your AI users the stakeholders can be the AI model itself right so um in especially in Industries like healthcare and finance right you know understanding this becomes very important and very useful um so again you know you have uh
tools for you know metadata management or data cataloges know you have know um or data cataloges know you have know um Apache Atlas you have you know kibra um Apache Atlas you have you know kibra um there's another one um Amud Zen or
something right so there's another one that actually helps in that metadata and capturing that data lineage um for the data that is being fed into your model right and then of course lastly uh you have you know data Ops again standard
you know you can build a model you can have data associated with that but if there is no continuous deployment into the production it remains as a proof of concept itself which 80% of the companies end up doing anyways right so
how do you make sure that you are automating your data pipelines the data that feeds those models having them if there are changes being made uh from a there are changes being made uh from a data standpoint or from a uh tables or
making sure that you're continuously deploying them then of course you know cicd for model runs I think data braks um has this very nice concept wherein you can track model registry know you have feature stores where you can use
features run the models track the model registries you know uh look at them from how do you make sure that you're running that cicd for the model runs themsel and talked about we always talk about concept drift wherein
the inherent nature of the features themselves change right so let's take an example right you know say for example a company is uh say Reliance again for the lack of just as an example right say Reliance had uh their own mom and pop
stores back in the day they were generating data right now suddenly they came up with a website with an online store right so the inherent nature of the data itself changed wherein you know the channel with which they are selling
has changed right so now this channel becomes an important feature earlier channel was just offline now channel has two values online and offline right so that can actually impact your model performance so how do you make sure that
that inherent change in the nature of that feature is being incorporated as it from a data standpoint into your AI model makes it very important right so when I say concept features can also drift right so making sure that your
Contin continuously um checking for your models to make sure that they are working on the latest data and continuously making sure that any new data that is coming through is being fed into a model to improve its accuracy or
improve its performance I think becomes very important right so to summarize you know it helps building the foundation of AI data you know ensures foundation of AI data you know ensures that the model or the data is AI ready
or the AI is data ready and then of course the process makees sure that you continuous Improvement you know for for your data so things that sometimes you know we kind of tend to push it to the back end
but you know very important from an Enterprise structure standpoint so thank you you know that's me you know if you have any questions I think do we have time yeah if you have any questions we have a couple of minutes
everyone and again thanks code basics for the opportunity thank you so
hour where you can enjoy high tea as well as networking first break we do have two interesting sessions I request everybody to be here by 4:30 thank you so much I know it is overloaded you need
overloaded you need refreshment sorry five I'm going back refreshment sorry five I'm going back okay 5:00 be here by
hello Basics is to make complex topics simple and for
in their place because we are starting our session
have you collected your t-shirts guys have you collected your the session you can go to the code Basics and then you can get
we have the next session on building AI applications with model context protocol applications with model context protocol and entic rag so um people uh here what
is the word that you listening most from the morning till now ai sorry data
right when I transited my career initially I was always wondering why initially I was always wondering why there are
not that easy but whenever you want to get into the any field maybe transition or anything you need to have one skills at least that can speak about you and that's what one skill has given me one opportunity in one of the companies and
then lately I have built the branches of you know learning other things I know there are so many people who are already into Transitions and mostly I find mechanical people how many are there mechanical here mechanical
engineering because in the panel also today I saw so many people like they were from mechanical engineering is it something that after taking me mechanical then you rethink about I need to go into data
analytics something like that anybody wants to share their
10 years of experience in building data engineering and artificial intelligence engineering and artificial intelligence based system currently she works as a generative AI product manager and solution architect pioneering several
use cases across critical functions of large pharmaceutical mnc's I welcome uh snea and I request you to please be on the stage for your knowledge sharing and I request everybody a round of applause
I request everybody a round of applause for SN thank you for the wonderful introduction um
today's topic is building gen applications with model context protocol applications with model context protocol and agentic rack I understand this is a uh this is a mixed crowd so if you've already playing been playing in the Gen
Rag and ecosystem this present may be super easy to follow uh but if not maybe I'll I'll give you some pointers for you to explore more further after the to explore more further after the session so either ways uh maybe I'll try
to level the play playing field for everyone for the next 20 everyone for the next 20 minutes so uh
around the AI geni Community maybe in the last one month or so so who came up with it it's the company called anthropic and they came up with this protocol in November 2024 so anthropic is a AI based company
based out of San Francisco in the United States uh if You' have heard this llm States uh if You' have heard this llm claw so it came up it it's an claw so it came up it it's an llm designed by or trained
so what is model context protocol and why is the big de so basically model why is the big de so basically model context protocol it standardizes how gen applications provide context to llms so moving further if you see this
diagram right so on the left reading this diagram from the left to the right llm by itself is very limited in its knowledge for example it can only answer questions based on the data it is Stained on let's say for example it was
trained until 202 4 it can only answer questions until 2024 but it cannot catter to your Enterprise use use case specific questions right but so moving to the middle middle layer right so in this
case llm is connected to different data sources like maybe a vector DB or a sources like maybe a vector DB or a custom application or a web search tool right so giving you the responses in this case llm responses are grounded in
the truth or the data that you provide still better you know you can still it from A to B moving from left to right you have you have enhanced its
capabilities making the llms actually respond to questions based on or grounded in your truth in your Enterprise truth right data stores documents Etc moving to the right side uh from the middle to the right
what's Happening Here is if you design applications like in the middle every time you swap out or change a data store you have to change the AI logic that the way the llm connects to the data
source what model context protocol does is standardizes all of this making AI application Arch architecture design much more easier because in this in the
right case rightmost case with model context protocol LM is only talking to context protocol LM is only talking to the mCP server right which exposes Vector databases or custom apps
exposes Vector databases or custom apps or web API maybe Bing Sur apis etc etc so as it's easy to swap in and swap out the tools that you want your llm to get the tools that you want your llm to get context out
model context protocol what are some of the wins is it streamlines the architecture make because llm only has to interact with mCP mCP then integrates with the rest of the
world that means highly interoperable and context Rich AI chat boards or geni and context Rich AI chat boards or geni chat boards or
logic from on the data source of course so the second major win you all know this like uh context window length is one of the limitations of AI models so every different AI model has a specific context length window in which it it can
respond to your questions within the context that you provide so so for some of the popular models out there this is a context length for example the latest a context length for example the latest gbt 40 has about 128k token length how
you can approximate it it is about one word is about 1.3 tokens so if it is word is about 1.3 tokens so if it is 128k tokens it's about 96,000 Words which is about uh maybe in English language it's about 350 Pages book think
of it that way right but if you upload multiple 356 Pages book it cannot it cannot take context out of both the documents and respond to your questions
right and let's say your conversation chat history is longer then also it loses context and again MC mCP comes to the rescue here
how I said I'll level the playing field when I said context when it's the amount of tokens in llm can process in one thread that's what it means thread that's what it means simultaneously so again so model context
simultaneously so again so model context protocol where it wins is it acts as a let's say If you think l as a brain mCP is a brain's extension it's a long-term contextual data mCP server can actually maintain
the long-term con Contex contextual data Beyond The 128k Tokens that llm internally has right so this is one this right architecture diagram is from anthropics website actually the model context protocol.com
actually the model context protocol.com so I really like this very simplistic architecture which can help visualize how m CP protocol how m CP protocol works so like I said mCP servers can get
the context from external data sources right let's say like vector databases maybe pine cone vv8 whatever your company uses or it may be um a document store or it may be you know web search any web search API it can not only
extract the information it's a two-way street it can also store the contextual information into the vector DBS let's say for example there are frequent queries that your user does on the application all of that can be stored as
an interim Vector embeddings on the M mCP server which can ex and provide context to the llm accordingly as and when
popular open-source uh mCP tool which is being used in the code interpreter applications right to code interpreter applications right to store code configurations documentation
here for you to like process this diagram so on the left is the mCP client diagram so on the left is the mCP client or the CLA desktop any of the ID tools making a prompt or a query it gets redirected to the right
query it gets redirected to the right mCP server which exposes the data mCP server which exposes the data apis and it's a two-way context exchange you can write to the data source and also retrieve from the data
source okay so and you all have heard this term probably if you're playing in this AI space for a while rag right rag yes retrieval augmented generation
so so uh I was asked this question if rag is going away is mCP going to replace rag so I'm going to come to it a little bit later but for the interest of everyone I just want to quickly run through what a
traditional rag is right it takes data from a data source let's say for example you upload a document uh the user the application developer has to configure the part ing strategy chunking strategy what's the
chunk size going to be like what's the augmented context is going to be like and how the embeddings are going to be done what embedding model to use for the domain specific data and how are you going to store it
Vector database are you going to store is it going to be a hybrid search search only all these are the decisions that an
scientist make when he's actually or uh helping build out the rag pipeline for a custom Enterprise use
Enterprise use case so again after the retrieval how to rank you know how to prioritize the results how to augment the context and is the and measuring the quality of the response right is is it really grounded
in the truth or is it hallucinating etc etc right so but what if the entire data in the back the data distribution the type of data Everything Changes than what it was developed for
it can happen right data is not data distribution nature of the data veracity volume doesn't remain all this same all the
doesn't remain all this same all the time so in that case traditional rag has some limitations because it's a one pass retrieval and let's say if your prompting is subpar if you give a subpar prompt it'll give you subpar results
because it won't retrieve from the vector database agentic rag came into picture so agentic rag came into picture where it adds an
additional intelligence layer to the retrieval augmented generation process so basically um this I'll I'll run through a couple of uh diagrams basically so this is uh from a from one
of the blogs online blogs so basically on the left side when the uh when the documents are ingested there's a document parsing summarization chunking
embeddings are stored in the vector DB right and then when from the CLA let's say cloud is your user interface where in the user prompts a query based on the in the user prompts a query based on the query mCP mCP server triggers the search
it decides whether it has to do a catalog narrow or broad search that's the agentic intelligence the Claud is llm which is deciding what kind of llm which is deciding what kind of search to do and then and then again llm
responds back with what search and the results based on the response again Claude or the llm agent take llm in distributes the results to each either you know web search or a chatbot
or a data analysis tool it could be anything so what's basically happening is that you're adding an additional layer of intelligence every step of the layer of intelligence every step of the way from analyzing the users prompt to
way from analyzing the users prompt to adding the context right context by by figuring out which is the right data source to direct and fetch the data from source to direct and fetch the data from and also measuring the quality of the
responses and the step by-step evidence inference of it everything is automated inference of it everything is automated by a agent agentic layer so it could be a single agent system or a multi-agent system so if you look at this diagram
from the left to the right what's happening here is that user sends a query and there's this master agent which kind of which kind of orchestrates its uh request to its
children agents and the children agents decide which tool to fetch it from is it from the vector database A or B is it going to be a web search is the is the users prompt requiring the latest web search results should I fetch it from my
slack is a file in this is a data available in my Enterprise team chat or like slack or Gmail all the decisions are made inherently underneath invisible to the user user when you use a multi-agent rag system so in essentially
what it does is over traditional rag it can help improve the response that you can help improve the response that you get out of the llms in your by providing llms more context Rich information so why mCP right all good
fancy intelligent system so where does mCP comes come into picture when you're dealing with a multi-agent drag
so so wa does it is it going to is mCP going to replace or is irrelevant what going to replace or is irrelevant what do you think in terms of agentic rag
probably when I was talking in the beginning mCP layer adds a more standardized way to to help swap in and swap out the tools the agent are going to interact with and and streamline the systems
further more let's say if you were to let's say you build a system with five AI agents right maybe one two let's say if it's a uh if it's a use case for customer support system you get the Open tickets and then you get the sentiments
or comment about the product from you know
uh of course a agent has an llm model underneath right that's what is making the decisions so it decides which probably server to direct the users request to and the second
is each of these different mCP servers let's say one mCP server to extract data from your Enterprise databases other mCP server for web search other mCP servers
to connect to other custom apps within your EOS system and the third is a vector database right this Vector database what it does is it can store the embeddings of the most frequent queries or user
inputs or some configuration settings that are retrieved very frequently and it can also be hosted on the mCP server on the mCP server RAM memory which can on the mCP server RAM memory which can increase the brain which can increase
the context length of the inherent llm and the fourth one is of course a client interface like your you know your chatbot web interface or Cloud desktop interface whatever is interface the client interface that the user is
client interface that the user is querying or prompting which uses model context protocol and agentic rag I think this is the data
flow how it would look like let's say it's a user query ingestion for example uh user query ingestion is you prompting chat jpt think of it like you prompting perplexity whatever right whatever the client
interfaces and maybe in a Enterprise use case if I were to take a customer support or customer success group and they open some your your customers open some tickets some issues with the product right and there's another
there's another customer feedback system where they put in you know the complaints they have about you know using the system or the feedback they have about using product so I I'm prompting this question
generate a report on open support tickets and include any recent related customer feedback okay that's the prompt that you have given to the llm and then what happens the llm inherent within the
what happens the llm inherent within the agent gets this and it starts planning where do I with server should I route this to should I route this to database mCP server or should I route this to feedback mCP server or should I route
mCP servers uh which are exposing data different data sources to the llm right so it agents is starting is using its brain it's planning once it decides on the server
planning once it decides on the server it decides on the flow mCP this these are the retrieval actions that are happening first the request gets sent to the the um where Open tickets so it gets the
10 Open tickets and then the 10 Open tickets gets sent to the feedback mCP server and there is a customer sentiment that's getting pulled in pulled in pulled into the context based on the Open tickets so there are two flows
Open tickets so there are two flows here and then now this context that we have gotten from these two mCP servers is being passed to the llm so it's it's also very important as a developer to also structure how we are passing this
llm context into the to the llm so probably you will say here are the relevant data this is the ticket data these are the feedback experts using these two data sources that I have retrieved please answer this
user query right the query is user query is generate a report on open support tickets and customer sentiment now what tickets and customer sentiment now what happens LM generates a
response based on the context so you can also have another agent evaluate the response as well to improve the quality of this and then you can also have this optional knowledge storage on the mCP server which stores the most common user
queries like you know if the give me feedback on so and so is if that's the most common query you know um store the embeddings of the prompt and the response in a mCP server so that it can be brought into context immediately
database server if you already have an agentic RG system right and you are bringing in model
context protocol newly these are some of the six steps that you need to look into one is collect and pre-process the data for retrieval right let's say you your data is document storage or your document store or your data is
maybe uh web search API sorry maybe your uh data is in big query whatever it is uh data is in big query whatever it is right um so prepare and collect the data em create the indexes on the vector database especially for the most
frequently queried Collections and then you set up the M mCP server deploy the server to interface server capabilities to the client as
soon as a the client receives a prompt it it needs to know what server it it needs to know what server capability it needs to tap server results right evaluating the response and then you configure the ml
response and then you configure the ml mCP and MCT setup so how do you improve for performance or accuracy some of the optimization strategies right so we've talked through the whole flow here so at
each step there's different things that you can actually look into in terms of optimizing um rag setup so caching and reuse where does it come into future like I've been saying right embeddings of frequent queries can
be on the mCP server or the documents that you're generating on the Fly for caching effectively and in multiple places also tuning the Vector database right so if the retrieval quality is
right so if the retrieval quality is poor the llm response lm's response poor the llm response lm's response quality is also going to be Poe so while a lot of emphasis is on tuning LM responses there should also be enough
emphasis on retrieval strategies and tuning the retrieval so how do you tune the retrieval right you based on the domain maybe you working with Healthcare know edtech data or maybe you're working with Finance data so choose the right
embedding model based on your domain to store the embeddings into the database and then at the point of retrieval tune the similarity threshold or topk if you make topk too low or too
high too low Also may not be may not work but small enough might work so look work but small enough might work so look into that and then also metadata filters right there are some people there are some rare keywords that the user might
search on in that case Vector search may be poor but keyword search will do be poor but keyword search will do really well so having the right metadata filters and also as you're maintaining the index of the most frequent query
collections also ensure if they're getting old just remove them and replace with the most recent ones you know don't keep the historic context not relevant now and the third is of course prompt engineering and agent instructions when
you're setting up any uh agentic system there's a system prompt that you configure right let's say for example if the users's prompt requires the latest web search results please get the latest web search results
but if you're not sure do not say you do not know okay that's my system prompt so it'll hand so it will also handle the failure gracefully right if it doesn't know it says it doesn't
know and then f short prompting especially with agents you can give a sample you know saying if you find this type of document do this if not do this give a sample prompt before you know try few short prompting
techniques and you know you can be very specific with the LM give me step by-step answers verify the retrieved answer before giving y out to the user all of that you can configure as either agents or system
agents or system prompts and the third is prompts and the third is have a good classifier which can route your users's query with the to to fetch data from the right data source right in
data sources and user can prompt anything so H so the agent which is deciding which the router has to be smart enough to decide which data to
fetch from based on the user prompt where how how the routing is happening and then again when you're using model context protocol ensure the mCP servers are optimized let's say for example MC if you notice that a lot of
example MC if you notice that a lot of queries are going to web search right the latest information which mostly needs 90% of your use cases are triggering the web search then look into using a more efficient API or scale up
the mCP server calling the exposing the API for better performance so next step is monitoring and ongoing tuning so when you're deploying a a agentic rag
application J application always know that you are um this is like a continuous evaluation system you're going to tune you're going to look into um how often a particular mCP server is used is it the database server that is
getting called often is it the web search is it the custom to apps MC mCP server which is getting called most often being able to notice that and how long each call is taking if some of the most used uh mCP servers are taking a
long time then you need to optimize for that if if it if there's an mCP server which is just occasionally being called then you can or not being called at all you can consider removing it from your flow okay and the success rate right in
an agentic system there's a lot of uh cases of failure so are the uh are the failures being handled gracefully that's another question so if it's not
answer on time what would you like to display to the user that will make them come back again to your application so basically I've been exploring this topic honestly since
March 1st 2025 uh there are the anthropic website has been like super helpful in me for me in understanding in addition to some of the high quality addition to some of the high quality blogs um written by these two authors um
and I'm going to share it on my LinkedIn you can actually follow me here um and I you can actually follow me here um and I post uh things on data Engineering in AI post uh things on data Engineering in AI uh latest readings that I do my my
perspectives etc etc and that's about it so um I'll be here for some more time if you want to catch up and to catch up and talk thank you thank you snea thank you
talk thank you thank you snea thank you so much for that detail uh guys I have an announcement uh we have arranged ifar at 10th floor the people who are fasting they can proceed for ahead 10 minutes so that they should
not be getting that lift uh you know you can use lift take your time 10 or 15 minutes ahead you can reach there now 6:30 all 6:30 all right okay now I will call upon him and
right okay now I will call upon him and Padmini for
an end note now and uh do not move we are going to play something very interesting so what all we did since morning we have captured that that to you followed by that there is going to be an end note and after that
there's going to be a lot of networking all right so the video should be playing all right so the video should be playing now in a minute
oh [Music]
[Music] C
oh now [Music]
I [Music]
there [Music]
you [Music]
captured in the video oh all of video oh all of you oh we are so connected and we see ourselves in each other is it so that's so great
so all these things were great right like uh we have had this uh event and uh things today what was the most interesting session for you today what
interesting session for you today what was the most interesting session have learned today today sorry final discussion okay which one
the data analytics or the data science data analytics okay glad I was there thank you okay so yeah it was uh it was all amazing uh you know to have this sessions so I personally learned a
variety of topics again from Agents to mCP then we covered from um we covered and a lot of other things and we also looked into the business side of things that should be the key takeaway for you uh can someone give me the pointer
okay but how many of you remember this picture from the morning everyone right so again for all you know all this the entire day could be an hype right could be an hype you don't know I don't want you to trust
anything blindly go home do your research and learn to suck the reality research and learn to suck the reality out of what hype and misinformation so tomorrow morning when you're sipping the coffee the first sip should be what the
reality not misinformation for that you need to use a straw consider this as a suck the reality from hype and misinformation all right so um see I'm I
can't stop talking about how this AI is impacting Healthcare particularly I was very interesting how this Google's Deep Mind project all this Alpha fold and all this what ai ai
what ai ai ai how many of you got super bugged of listening to this saturated abrupt entry I'm here to
entry I'm here to give an end note let me briefly give an end note let me briefly introduce myself my name is padmin I'm a marketeer by choice and profession in a nutshell what we do
is make good products great all the Brilliant Minds here your Brilliance has to be seen by someone right that is why we exist we have to Market your wonderful products your brain
brain so today I was just so today I was just wondering that
because I've been uh listening to terms like uh Azure or snowflake T Lake for me I
think fabric yeah algorithm mCP I agents agents algorithm actually sounds like a agents algorithm actually sounds like a fancy dance move to me you know I
totally it's a bouncer but few of the sessions I did get a sense of it so sorry for the slight she got so confused she's mov moving backwards now
confused she's mov moving backwards now oh so I was just giving him an example cloud computing for instance it looks something like a big fluffy Cloud full of uh tiny computers raining
down my imaginative head thinks of it like that snowflake also yeah how can like that snowflake also yeah how can snowflake be something related to data I want to learn all this someday I I'm very interested so you are you the only
like you did anyone feel like that hearing this words this fancy terms okay any more people who feel like this fancy computers falling from the cloud no in my team are full of data and N
no in my team are full of data and N sometimes in meetings also I get so you know out of the place I feel what are these guys doing okay I have to ultimately marketing Market what they are making great product your boot our
boot camps are the best so I have to some but still I have to know what it is so my next step is to actually learn some of these terminologies so I was talking about these this imaginary you know I I took help of a I though I'm not
that Savvy I somehow try to give a prompt this is how it looks looks for me okay so I'm sure all of you had a wonderful experience today Geniuses came
wonderful experience today Geniuses came gave you all the Gan but we are here always I'm always trying to give a human touch to things so that reminds me we actually curated a cute little bag many of you also asked
me so you all got your bag right the welcome kit how many of you welcome kit how many of you found an envelop that set secret so many of you asked me yeah how
secret so many of you asked me yeah how many of you opened it right hey you didn't get that uh Lottery some people got like uh some people got like uh $1,000 yeah some people got 1,000 some
people got 200 rup some people got one rupee some people got caught yeah you're here that's not true I think some some mess from our team yeah something got
messed so what made you do that what made you like CH what is happening correct you're going to get it's already up there Curiosity our brains are always
the human mind like I said we are always curious I always keep telling this somebody was curious so we landed on the moon what is there on the moon that's curiosity leads to discoveries curiosity leads to new
inventions our mind has to always be curious but can I ask you something with this AI everything AI can something with this AI everything AI can predict future Trends AI can process
data AI can probably dance better than you if I program that robot it can dance even cooler but can AI give you a subjective Human Experience can it give
you you know curiosity is what made you will AI ever curious no maybe not we have not reached that stage yet and we'll never do be able to get to the human side of things so let me just get this quick be with me for 2
minutes more just sit relax get cozy get to know your neighbors at the count of three I want all of you to be very
three I want all of you to be very honest and shut your eyes for just 2 minutes and just stay still be there with me I promise you'll thank me later for this from wherever you are give yourself
this from wherever you are give yourself a moment and take three deep breaths so inhaling and feeling the opening and feeling the opening and expansion and exhaling completely
releasing what isn't needed perhaps clench muscles in the jaw shoulders and
and simply allow yourself to step back into open yourself to step back into open awareness being with whatever arises and that may be the breath
sounds images or physical Sensations allowing these experiences to pass through like
clouds and behind it all is your awareness like the sky your awareness is open fast with the space to hold whatever fast with the space to hold whatever comes and goes
and as you're ready can take one more conscious conscious breath and then
feel little at least good little little good yeah
closed your eyes no okay that's good all right but me you know what you did try to give them a human experience
did try to give them a human experience what you did to me you made me something human I'm going to recognize and thank the
people who made this show happen these are the main characters of the show I'd like to call them on the stage that's the code Basics team please come on to the stage give it to them guys they are everything who did this and made this
many of them haven't slept yesterday and still I think their body is dead they're running with their soul and and I can't you know like uh you know this this couldn't have not happened without them you know doing this and but if I think
more about it this could have not happened even without you you being the happened even without you you being the audience so let's all take a b to the audience so let's all take a b to the audience like thank you thank you thank
you you have been really patient with us you have been uh you know like really supportive and is it just you no we want to thank our sponsors as well who kind of helped us in organizing this event in a better way and the speakers who shared
a better way and the speakers who shared a lot of knowledge right and we want to a lot of knowledge right and we want to thank our host networking Stars volunteers and everyone and there's a little video we have made to dedicate
little video we have made to dedicate our gratitude for all of them and freedom [Music] for being here we going to do not forget the human side of things start making
meaningful connections and it's a time to network for the sake of humanity to network for the sake of humanity thank you
[Music] know
[Music] clap on
they have been with us since yesterday night they have been packing all the night they have been packing all the gifts and everything office yesterday and they have been packing something a present a present
for all of you for your presence this will be collected this can be collected kind of a coupon please redeem it at the ground floor from 6:30 onwards right now building those meaningful connections and yeah we'll take more photos we'll do
more networking and for people who are breaking fast there is an ifar available at 10th floor and yeah it's it's a great day I look forward to see you outside as day I look forward to see you outside as well thank you