Let’s build K.I.T.T. with JavaScript
summary
A K.I.T.T.-like agent runs entirely in the browser using JavaScript, combining transforms.js with Whisper for ASR, Kakoro for TTS, local LLMs (Quen 3 / Gemma nano), WebGPU/WebAssembly for acceleration, and WebBluetooth to drive a Lego car via an Arduino Nano RP2040 exposing a two-byte characteristic. The pipeline indexes 6,275 scenes, creates 384-d embeddings, streams ~54 tokens/sec, processes 3.1 million input tokens into 303k outputs for RAG and tool-calling; the demo works offline though occasional crashes occur.
AI-written summary · about
chapters (18)
- 0:00K.I.T.T. in JavaScript
- 1:06Speaker and project
- 5:24Hardware and ML frameworks
- 7:36Transformers.js for ML
- 9:46Speech with Whisper and Kakoro
- 11:34How LLMs generate text
- 13:22Small LLMs in browser
- 16:14Streaming LLM responses
- 19:14Prompt engineering for Kit
- 22:50Memory from TV scripts
- 24:34Vector embeddings for search
- 27:44Summarizing scenes with LLMs
- 29:09Tool calling and RAG
- 31:54Kit’s memory and reasoning
- 35:29Lego body with Web Bluetooth
- 38:16Web Bluetooth API control
- 38:58Kit’s agentic workflow
- 41:06Why build K.I.T.T.
related talks
43mLet’s build K.I.T.T. with JavaScript
30mLet’s build K.I.T.T. with JavaScript
33mLet’s Build K.I.T.T. with JavaScript
40mLet’s build K I T T with JavaScript
30mLook ma, no hands! Multimodal AI-Agents in the browser
29mClient side Web AI Agents Building smarter user experiences for a future age
22mClient side Web AI Agents: Building smarter user experiences for future agentic ...
22mBoring but Effective: WebAI That Ships