[00:02] [music] that are inspired by production grade AI systems. Systems that uses rag, multimodal inputs, agentic architecture, and real-world integrations. In this video, I'm going to share such five projects that you can build [music] in
[00:18] year 2026. These are not random projects that I have picked from internet. These are based on the work [music] that we do in our AI consultancy company AtliQ Technologies, where we sell to the clients based in US and UAE. I will be
[00:33] sharing a PDF with project details, tech stack, resources, and some of these [music] projects are available as guided video tutorials, as well, to make your learning journey easy. These are the five projects that we will
[00:48] be discussing. Let's start with our first project, which is rag with role-based access control, guardrails, and monitoring. In our company AtliQ, majority of the projects that we
[01:02] majority of the projects that we implement in AI are rag projects. Okay, industries. So, here, what you will do is you will have some internal knowledge in any company. It can be PDF, Excel files. You
[01:20] know, knowledge can be present in variety of formats. Now, the LLMs do not have information on your company's private data. Also, LLMs are trained uh till certain date, so they have a knowledge cutoff. So, if you ask any
[01:34] latest question, the LLM itself would not be able to answer it. So, it needs external tools, it needs external knowledge, and that can be achieved by retrieval augmented generation, okay? So, here you are building an internal
[01:50] chatbot for a company, and you will limit the access uh using RBAC, okay? So, let me just show you the data set. So, we ran this
[02:02] resume project challenge where we have given all the data, okay? So, once you way. Once you enroll, you will get this download files button. Once you download those files, you will get this kind of
[02:18] data set, okay? And there, you will have data for engineering department, you will have data for finance department. Now, using role-based access control, if department, then only they get access to finance data. If a person is working in
[02:34] engineering, then only they get access to engineering data. If a person is CEO or let's say CTO, then they get access to all the data. So, you have to code your logic using uh metadata filtering in vector
[02:47] database, okay? And we have given the problem statement and everything. So, for this project, you have all the ingredients, data sets, and everything that you need. Uh and then, once you build the project, you need to set up
[03:00] important, folks. You need to protect sensitive and PII data, handle out-of-scope questions. And what I mean by out-of-scope question is,
[03:12] here is the example of Chipotle's real customer service chatbot where somebody asked a question on how to reverse a linked list, and it answered that. This chatbot should help you with the orders, not with coding, okay? So, you should
[03:26] build your AI project such that it doesn't do mistakes like this. Then comes deployment into Azure, okay? You can deploy it in either Azure, AWS, etc.
[03:38] where you will set up monitoring and evaluation, and you will also do cost monitoring so that you make sure your token cost doesn't go out of your budget, okay? For technical stack, you will use LangChain or Docling along with
[03:54] Python for this project. Docling because it allows you to extract the information from unstructured sources such as PDF files. So, this is a a library by
[04:07] IBM. So, it's an important one. And the LangChain is a framework that is very popular in the AI industry. So, you will use LangChain to write your rag code.
[04:19] For vector database, you will use either Quadrant, Chroma DB, Milvus, etc. Vector databases will store the embeddings of the text and you can do semantic search using it. For front end, you can use Streamlit. Or even better, with Vibe
[04:33] coding, you can even build a React web application. You don't need to know React coding. And for LLM, you will use GPT-OSS, Llama, etc. inside Groq Cloud.
[04:45] Groq Cloud allows you to call LLM for free. Okay? So, this is perfect for your learning purpose. You can access all these models and you don't have to pay any money. And for cloud, one of these three. For evaluation, you can use
[05:01] Ragas. So, let me just show you. So, Ragas is a library that allows you to uh evaluate it. And then, you can use LangSmith for monitoring. So, LangSmith is part of the uh LangChain ecosystem and it will allow
[05:17] you to observe, evaluate, and deploy your agents or even your regular AI project. Okay? So, this is how your rag pipeline is going to look like. I already showed you the page that we have for where you can download the dataset,
[05:32] etc. And then, I'm going to link some YouTube videos and some uh useful links that will help you build this project. The second project is going to be the voice agent. Now, this is the project that we have built in Atlic for one of
[05:48] our clients in Middle East and whatever I'm telling you is based on the real experience that we had. So, here many companies want to automate their call. Let's say you are calling customer service and they want to automate that
[06:01] call. So, what happens is voice agent needs to be designed in a way that it should be able to listen to a caller, convert speech to text, understand the customer's intent using LLM, retrieve
[06:16] relevant information from company using knowledge base, so using rag, and then respond back in a natural speech using text to speech. Okay, so this is how the pipeline looks like and then you should be able to answer FAQs, check
[06:32] order status, schedule appointments, etc. And if there is a complex problem, you can also route that call to a human agent. Here, you can use either an open-source platform such as LiveKit or you can use a commercial
[06:48] platform such as Eleven Labs. So, in the project that we worked for our client, we used both the approaches actually. So, Eleven Lab was one solution, but example, in Middle East, they have certain constraints on what kind of
[07:03] cloud solutions that they can use. So, then we went ahead and implemented this with a LiveKit. So, LiveKit is a open-source platform. You can host it
[07:15] machine. Eleven Labs is more commercial, easy to use. Okay, so these are alternatives. So, let me show you how Eleven Labs works. So, Eleven Labs provides you 10,000 free credits, so you can use it
[07:32] as your learning project. And once you sign up, you can go to Eleven Agents. And here, see, you can create your voice agent. So, you can go here.
[07:44] You can browse through your templates and let's say e-commerce customer experience chatbot, okay? So, you will use the template and then you will do your customization. So, here you can specify the personality, your basic
[07:58] system prompt, okay? You create an agent and in this agent, you will load the voices, language, which LLM you want to use. Okay, and once uh this agent is set up, you will work on a workflow. So, here
[08:13] they have given you a workflow. You can customize it using this uh UI editor. You can also specify a knowledge base. So, when someone is calling a customer service, the customer service agent usually will have FAQs. They will have
[08:28] company information, information on their services and products. You can their services and products. You can attach all of that here in a knowledge base and you can configure your rag here, okay?
[08:40] And then, you can also attach certain tools. So, let's say you are working with Verizon customer care or let's say Reliance Jio customer care. You have a problem with your internet. Now, uh you need to schedule an appointment.
[08:53] So, in that case, uh here you can add a tool which can schedule an appointment and the uh technician, you know, you can have access to the technician calendar and this can uh schedule the appointment using this particular tool.
[09:09] You can end conversation, you can detect language. Tools are the functions and utilities that agent will use to uh perform certain task, okay? You can have MCPs as well, okay? So, you can play with it uh
[09:24] with it uh and once uh the conversation ends, it will be logged, okay? Here, see. And you can monitor them for the quality. So, this is a a great platform if you
[09:37] So, this is a a great platform if you want to build any voice agent. Now, want to build any voice agent. Now, here I'm going to show you the technical architecture. So, by the way, when you make a phone call, see user makes a
[09:49] phone call from here, it will go to telephony provider. So, that is just show you here. Okay, so in that way, you can just go to phone numbers and import a phone number either from
[10:02] Twilio or SIP. So, Twilio is a service which provides you this calling support, okay? So, you can get a phone number, like a digital phone number. You can import that. And then, you can give that to an agent. So, what happens is
[10:17] user makes a phone call, let's say you have given a Twilio phone number, it will go to the telephony provider, which is Twilio, and Twilio will send it to is Twilio, and Twilio will send it to the agent in ElevenLabs, okay? So, in
[10:31] ElevenLabs, this is real-time media server, you will convert that to speech to text. So, for your agent here, in this agent, you can add this speech-to-text. And then, you uh basically, you convert
[10:46] this speech to text so that you can make a rag call, and you can get the better uh outcome, better answer for the query that is being asked. And at every point,
[10:58] folks, see, you are doing observing and infrastructure, okay? You are logging, you're monitoring, you're handling errors, and so on. When you convert uh errors, and so on. When you convert uh it from speech to text, that text can be
[11:11] it from speech to text, that text can be given to your LLM, okay? So, right now, given to your LLM, okay? So, right now, my LLM is a Quen model. You see this Quen model here. And this model will use rag. It will do
[11:24] tool calling, it will use knowledge, memory, guardrails, all of that to produce an answer, okay? So, you see all this business tools here. So, you see my AI, database, everything. And once the answer is generated, you convert text to
[11:41] speech and you return response back to the user. Now, architecture looks good, but in making it production ready, you will face number of challenges. For example, tone detection, interruption handling, silence management, and there
[11:56] are techniques to deal with it. There is something called VAD, voice activity detection, where it can tell you, based on the audio track, that this is speech and this is non-speech. And if you see non-speech for, let's say, 500 ms, then
[12:12] that you can mark it as a silence. Okay, user has stopped speaking. Same thing you can do to detect the tones, you know? You can detect the tones between two people who are talking. So, all these
[12:27] production challenges are going to be important to handle in this project. Okay? For experience, you want to give natural voice. When there is a silence, maybe you can show some typing sound in between, so
[12:41] that you're mimicking or simulating real-life situation. Okay? Economics is something you have to be very careful about. Model routing, token efficiency. Model routing means for simple intent detection, you can use
[12:55] low-cost model. For complex question, you can use high-cost thinking model, and so on. Okay? We have mentioned the tech stack here. You can take a look. For speech-to-text, you can use providers such as Deepgram, Cartesia,
[13:10] and for text-to-speech, same along with Eleven Labs. Okay? So, that's your voice agent project. And we have given some of the resources here. The third project is
[13:23] going to be multi-agent coding assistant. Okay? So, AI agents are very popular nowadays, and here you are making a coding assistant similar to Cloud Code. You give a prompt, and it will build an application for you. Okay?
[13:40] So, the way it will work is you will use LangGraph, which is a multi-agent uh framework. It's part of the LangChain ecosystem. And you start, you plan your
[13:53] software project. Let's say you want to build a calculator web app. You will plan, you will architect, then you will code, and there will be loop. It will just keep on coding, and there will be end. Okay?
[14:05] Now, the good thing about this project is it is available on YouTube as a guided tutorial. It is one of the videos on Code Basics channel. You can watch it on Code Basics channel. You can watch it for free. And here in this uh 1-hour
[14:17] project, we have explained the entire project along with architecture, along project along with architecture, along with coding, along with data set. I I you are just creating uh an app here,
[14:30] everything. We have given all the code. Okay? So, please take a look at it. And then we have given some exercises also, so you can make some modification. The so you can make some modification. The fourth project is multimodal AI
[14:44] assistant. Multimodal means you can handle not just text, but images as So, let's say you have uh an account in some bank. Uh let's say credit card, and you are
[14:58] working as an AI engineer in that credit card company, and you are building this chatbot where customer can upload their financial statement, and they can say, "Okay, what is this EC2 charge?" And this see, they are asking, "Why was this
[15:11] charge detected?" This is our AI financial assistant. And it will tell financial assistant. And it will tell you that this charge is due to so-and-so reason. Okay? So, here you are uploading an image. How do you handle image? Okay?
[15:25] an image. How do you handle image? Okay? So, for image, you can use some image specific models which you have it available in Grok Cloud. So, you see there is this vision model that you can use. You can use some other platform,
[15:38] some paid platforms as well. And this is going to be the flow. Okay? You upload a credit card statement, you say, "Okay, what is this charge about?" and it should tell you. You will use Doc Ling because you want
[15:50] to extract information from your image, PDF, whatever. For vector database, you can use Qdrant, Chroma DB, Milvus. Front end can be Streamlit or even React app. end can be Streamlit or even React app. And this is a rest of the technical
[16:05] stack. Okay? The fifth project is going to be text classification using regular to be text classification using regular expression, statistical ML, and LLM. Now, as an AI engineer, it's not like you are working on generative AI and LLM
[16:19] projects all the time. You need to have knowledge on statistical ML as well. So, here in this project, you are classifying programming logs. Okay? So, let's say you have all these different programming logs. You want to classify
[16:32] them as security alert, resource uses, workflow, error, etc. so that you can set up some kind of alert system which will notify your technical staff. And
[16:44] worked for our clients in Atli. It was a different domain, but because of NDA, I cannot share more details. So, I'm giving an example of a programming logs So, the architecture was that if you have static patterns, you can classify
[17:01] it using regular expression. And let's say the static pattern is always predictable. In that case, you can try BERT BERT encoding and then
[17:13] logistic regression which is statistical modeling. And then for even more complex patterns, you use LLM. Okay, so this is the real architecture that we did for one of our clients, and clients was very happy because
[17:27] it was not just saving them cost on the LLM calls, but it was pretty fast and accurate. So, as an AI engineer, you will end up building this kind of hybrid architecture, okay? So, this is what you're going to build, and this is how
[17:43] the technical stack looks like. And this project is also available as a guided project on our YouTube channel. You can see it's like 1 hour and 22 minutes, where we have given everything, code, everything, okay? And you can make
[18:00] code, everything, okay? And you can make some customization as well. Now, if you're looking for additional data set for text classification, okay? We have given the data set, but you can work on a different data set, so
[18:15] you can get the data sets from Kaggle, you ask ChatGPT, and it will provide you the help. So, by working on these five projects, you are covering rag with guardrails and monitoring, you're covering the voice agent scenarios, you
[18:31] are covering multi-agent architecture, multimodal scenarios, and also multimodal scenarios, and also statistical ML and some rule-based AI as well. So, this gives you a lot of diversity, diverse knowledge, which can
[18:45] help you in your resume, which can help you learn AI in a holistic manner. Now, folks, once you have built all these projects, make a presentation and post it on LinkedIn so that it increases your online credibility. For our RPC
[19:00] for example, these people were the winners, and if you look at their LinkedIn post, they have not only posted their code on GitHub, but they created
[19:12] their code on GitHub, but they created this kind of a video presentation as if you are presenting to your business stakeholders, okay? So, this kind of stakeholders, okay? So, this kind of presentation and posting it on LinkedIn
[19:24] will help you increase your online credibility. I hope you found this video to be useful. If you have any questions, you can post in the comment box below and share it with your friends who are learning AI. Thank you for watching.
[19:39] learning AI. Thank you for watching. >> [music]