AI Summary
The video demonstrates a fully local, private AI assistant called Jarvis, built on free tools like Ollama and Kokoro, that runs entirely on a Mac without internet, accounts, or API keys. It showcases a dashboard with a neural core for model management, a reactor for voice interaction, and a galaxy for semantic note retrieval, along with memory and tool-use capabilities.
Chapters
The AI assistant runs entirely on a Mac, with no internet, account, API keys, or cloud. The brain is downloaded once and owned by the user.
To keep personal data like plans, client notes, and routines private, unlike most AI chat apps that send data to servers. The system runs on Ollama, a free tool for local AI models.
The dashboard has three parts: the neural core (left) showing downloaded models, the reactor (middle) for voice interaction, and the galaxy (right) for the second brain.
Models include Gemma 3 (4B), Qwen 2.5 (7B), and GPT-OSS (20B). Larger models are smarter but slower and use more memory. Blue bars indicate easy fit, orange bars indicate heavy memory use.
Jarvis uses Qwen 2.5 because it supports tools, enabling actions like setting timers or writing files. Gemma 3 lacks tool support. Docker's tests found Qwen 2.5 fast and reliable for tool use.
The reactor ring changes color: cyan for waiting, amber for thinking, red for listening. The voice uses Kokoro, an open-source TTS with 82M parameters, 54 voices, and 8 languages. It speaks sentence-by-sentence for responsiveness.
The bottom-left panel shows CPU, memory, and token speed. On the Mac, Qwen writes 20-30 tokens per second, faster than reading aloud.
Every dot is a plain-text note. A local model (MXBai Embed Large) converts notes into embeddings and places them in 3D space, clustering similar topics. Asking a question lights up the relevant note and draws a line to it.
Jarvis can remember facts, saving them to a file that persists across reboots. For example, 'Remember that I record videos at 2 p.m.' is stored and recalled later.
Users can store notes about member questions or rules, and Jarvis answers from those notes, keeping data private and local.
The video presents a practical, private AI assistant that runs locally, offering tool use, memory, and semantic search, with a strong emphasis on data privacy and customization. It positions early adopters of such local AI setups as ahead of the curve.
Mentioned in this Video
Tutorial Checklist
💡 Key Takeaways
Privacy-first AI
Highlights the core value proposition: keeping personal data local, contrasting with cloud-based AI.
01:05Tool-use model selection
Explains why Qwen 2.5 is chosen over larger models, emphasizing the importance of tool support over raw size.
02:17Semantic note retrieval
Demonstrates a practical implementation of a second brain using embeddings, a key AI technique.
05:36Persistent memory
Shows how memory can be implemented locally, a feature often associated with cloud AI.
07:06Full Transcript
[00:04] I asked Jarvis one question. Jarvis, where are you running right now? He where are you running right now? He said, I am running 100% offline on your Mac, sir. That's my own AI assistant and it's part of my agent OS. It talks, it
[00:19] remembers, it uses tools, and it reads my notes. None of it touches the internet. There's no account, no API keys, and no cloud. You download the brain once and it's yours. That's what I mean by free forever. In this video,
[00:35] I'll show you the whole thing working, the prompts I use with it, and how to get it running on your own computer in about 2 minutes. Stick around for the part where I ask a question and a line of light shoots across a 3D galaxy to
[00:48] find the answer. That one surprised me the first time. So, why build this? I wanted Jarvis from the films, but private. your plans, your client notes, your daily routine, your goals. Most AI chat apps send all of that to a server
[01:05] somewhere. I wanted the private stuff to stay on my own machine. So, the whole stay on my own machine. So, the whole thing runs on a free tool called Olama. Ola lets you run AI models right on your own computer. You download a model and
[01:18] it lives on your hard drive like any other file. And my machine is a normal other file. And my machine is a normal one, a Mac with an M4 chip and 16 gigs of memory. The project is tiny, too. One server file and one web page. Okay, the
[01:33] dashboard has three parts. On the left is the neural core. In the middle is the reactor. On the right is the galaxy, which is the second brain. I'll save the galaxy for last because you need the first two for it to make sense. The
[01:47] neural core shows every model I've downloaded. Gemma 3 at 4 billion from Google, Quen 2.5 at 7 billion from Alibaba, and GPT OSS at 20 billion, the
[02:01] open model from Open AI. That billion number is the size of the brain. Bigger is usually smarter, but it's also slower and eats more memory. Each card shows the size, whether it can use tools, whether it can see images, and how much
[02:17] memory it fills. A blue bar means it fits easily. An orange bar means it's taking most of your machine. Here's a choice that matters more than it looks. I keep Jarvis on Quen 2.5. Not the biggest one. Why? Tools. A model with
[02:33] tools can actually do things. Set a timer, write a file, open an app. Without tools, it can only talk. Quick pause before I show you the next part. Inside the AI profit boardroom, you get the full agent OS zip file ready to
[02:49] install. That includes this Jarvis build, the second brain template, and the prompts I use with it every day. You don't build it from scratch. You download it and run it. We've also got a 30-day agent OS road map. Day one, you
[03:05] install it and pull your first models. Day two, you load your business notes into the second brain. Day three, you teach it how you work with the memory feature. We run four coaching calls every week where you can bring your
[03:19] agent OS setup and ask questions live. And lots of members are already running this exact stack. So if you get stuck, someone has usually fixed it already. Link in the description or go to aiprofitboardroom.com.
[03:34] Gemma 3 can't use tools. Google left that out of the whole Gemma 3 family. It's a nice model and it can even see pictures, but it can't do things. Quen 2.5 can. And I'm not the only one who picked it. Docker's team tested a bunch
[03:50] of local models on tool use. They found Quen 2.5 was fast enough for real-time use and still picked the right tool most of the time. To be fair, Quen 2.5 isn't
[04:02] the newest model out there. Newer ones use tools better. But on a small Mac, it's quick and it works. And swapping models later takes one click. Next up, the reactor in the middle. It's a glowing ring that changes color. Cyan
[04:18] means he's waiting. Amber means he's thinking. Red means he's listening. And when he talks, the rings move with his voice. That voice is Kakoro. It's open source and it runs on your Mac, too. Here's what surprised me. Kokuro is only
[04:35] Here's what surprised me. Kokuro is only 82 million parameters. The brain is 7 billion. So the voice is almost 100 times smaller than the brain and it still sounds natural. It comes with 54 voices in eight languages. I picked a
[04:50] British one because well it's Jarvis. When Kakoro came out, Victor Muster from Hugging Face asked, "Now that we have amazing open-source TTS with fast inference, what are you building?" This is what I built. One small trick makes
[05:06] it feel fast. Jarvis speaks each sentence as soon as it's written. He doesn't wait for the full answer. I asked, "How's the system holding up?" He said, "The system is operating efficiently with adequate resources." He
[05:21] got that from the telemetry panel in the bottom left. That's your computer's health stats, CPU, memory, and how fast the words come out. On my Mac, Quen writes about 20 to 30 tokens a second. which is faster than you can read out
[05:36] loud. If Jarvis ever feels slow, you look here and you'll see why. Okay, now look here and you'll see why. Okay, now the galaxy. Every dot is one note on my computer, plain text files. Each color is a folder. Here's how it works in
[05:52] is a folder. Here's how it works in plain words. Jarvis reads every note. a small local model called MX by embed large turns each note into a list of numbers that describe what it's about. Then he places the notes in 3D space.
[06:07] Notes on the same topic end up close together. It's like a library where the books shelf themselves by topic. And here's the moment I promised you. I here's the moment I promised you. I asked, "What are my Q4 goals?" The Q4
[06:20] goals note lit up. A line shot out from the center to that dot. and he answered from that note. I never told him which file to open. He found it himself. That's the part that changed how I use AI. Normal chat apps know nothing about
[06:35] you until you paste it in every single time. Jarvis already has it. Here's how I'd use it for the boardroom. Drop in a note with the questions new members ask in their first week. Then ask, "What do new boardroom members get stuck on
[06:50] first?" He answers from your own notes. Is it perfect? No, it's a 7 billion model on a laptop. Sometimes it grabs the wrong note. A big cloud model still wins on hard thinking, but for what's in my own notes, it's very good and your
[07:06] notes never leave your desk. Next is memory. I said, "Remember that I record memory. I said, "Remember that I record videos at 2 p.m." He said, "Noted, sir. That fact gets saved to a file and shows up in the galaxy as a new red dot. turn
[07:21] the computer off and back on and it's still there. I asked, "When do I record videos?" He said, "You record videos at 200 p.m., sir." For the boardroom, you'd say, "Remember that every new AI profit
[07:35] boardroom member should get a welcome message pointing them to the 30-day road map." Now, he knows that rule every time you work on new member stuff. There's a owners haven't figured this out. They're still pasting the same stuff into chat
[07:51] apps every day. That window won't stay open forever. The people who set up their notes, memory, and routines now will be way ahead. If you want my full setup, it's inside the AI profit boardroom. You can download the full
[08:05] agent OS zip file ready to install with Jarvis local core, the second brain template, and the memory setup you just saw. You get daily tutorials showing how to add new tools and models to your agent OS as they come out. Four coaching
[08:21] calls every week where you can share your screen and get help with your own Jarvis. A prompt library with the prompts I use for notes, timers, checklists and member onboarding and a member map so you can find other agent
[08:35] OS users near you. Link in the description or go to aiprofitboardroom.com. And if you want the full process, the And if you want the full process, the SOPs, and over 100 AI use cases like
[08:47] this one, join the AI success lab. It's our free community. Links in the comments and description. You'll get all the video notes from this episode, plus the video notes from this episode, plus access to over members.