Speaker
Description
Every AI tutorial starts the same way: sign up for an API key, add a credit card, spin up a server. For students and independent developers, that's a barrier before you write a single line of code.
This workshop skips all of that. Participants build a working AI agent inside one HTML file: streaming chat, custom persona, multi-turn memory, and agentic tool-calling, all running in a browser tab. No API keys, no servers, no cloud bill. Chrome and a text editor is the entire setup.
The stack is open source: Google's Gemma 2 2B (4-bit quantized, ~1.5 GB), WebLLM from the MLC-AI project (an OpenAI-compatible chat completions API loaded from a single CDN import), and WebGPU for GPU-accelerated inference natively in the browser.
We build in four hands-on phases:
- Phase 1, The AI Loop (40 min): Load a quantized model via WebLLM, run a first chat completion, add token streaming, and wire multi-turn conversation memory. From a blank HTML file to a working chat in under an hour.
- Phase 2, Chat UI (20 min): Turn raw text output into a styled chat application with message bubbles, a scroll container, and a loading progress bar. CSS is provided so participants focus on wiring the JavaScript.
- Phase 3, Make It Yours (30 min): Pick a persona (fitness coach, travel guide, language tutor, game master) and customize the agent. Then a live demo of agentic tool-calling, where the model outputs structured markers like [TOOL wikipedia "Eiffel Tower"], which get parsed and executed against real APIs. Results feed back into the conversation loop, following the same observe-decide-act architecture used by production agent frameworks.
- Phase 4, Ship It (15 min): Save the HTML file, write a two-sentence description, submit.
Because the WebLLM API mirrors OpenAI's chat.completions interface, the patterns and code participants write here work identically with cloud-hosted models when they're ready to scale. But for this session, the whole thing runs offline once the model is cached.
What to bring: A laptop with Chrome (WebGPU enabled, check at chrome://gpu), VS Code with the Live Server extension, and any GPU, integrated is fine. Every phase includes a checkpoint file so nobody falls behind.
Any other info we should know?
Participants leave with a working AI agent they built themselves, running in the browser on their own laptop. No accounts created, no money spent, and a codebase they own completely.
The topic is technical. We write real JavaScript throughout the session. But the entry bar is low: basic HTML and JS familiarity is enough. The workshop is built for people who have never worked with WebGPU, on-device inference, or browser-based ML before.
The session is hands-on and interactive. Participants code along on their own laptops in real time. Every phase has a checkpoint file, so anyone who gets stuck can catch up instantly and keep going. The last 15 minutes are dedicated to customizing and submitting agents.
I've run this workshop before: as an evening virtual lab through AI House - https://luma.com/ai-n9i7. The session had participants with zero WebGPU experience building working agents within 90 minutes. I'm attaching the video recording from that previous session so you can see the format in action.
Session author's bio
Nishant Thakur is an AI/ML engineer with 15 years of industry experience and a Master of Data Science.
He currently works as a Senior Computer Scientist at Adobe, where he builds browser-native AI inference pipelines using WebGPU, ONNX, and WebAssembly. Lately, he worked on Code Mode, a technique that cuts LLM tool-calling token overhead by 85% while doubling tool-selection accuracy, which was presented at Adobe Bangalore Developer Day 2026.
| Social Media | https://x.com/thakur_nishantk |
|---|---|
| Agree to Privacy Policy and Notice | I agree |
| Please confirm that there are included headshots of all speakers in their profiles | Yes |
| In Person Attendance | In-person |
| Level of Difficulty | Beginner |