Let’s Build K.I.T.T. with JavaScript
summary
A browser K.I.T.T. is actually built: on-device LLMs (Gemma family, small 1B variants) run via transforms.js and WebLM with WebGPU/WebAssembly acceleration, Whisper + Kakoro for audio, and WebBluetooth to control a Lego car driven by an Arduino Nano RP2040. Memory uses a vector index of 6,275 scenes summarized by an LLM (3.1M input tokens → 303k output tokens). Demo runs on a MacBook at ~70 tok/s ingest and ~32 tok/s emit, works end-to-end, and requires app-side guards for duplicate function calls.
AI-written summary · about
chapters (26)
- 0:00Introduction
- 1:05What is K.I.T.T. and features
- 2:49Hardware and ML libraries
- 5:40Natural language input
- 7:23Adding intelligence with LLMs
- 9:51Demo: LLM responses
- 11:20Fine-tuning with prompts
- 12:45Demo: Custom personality
- 13:06Building K.I.T.T.’s memory
- 14:33Similarity search for recall
- 15:57Summarizing scenes
- 17:02Vector search setup
- 18:46Implementing function calls
- 20:10Live demo: Voice interaction
- 21:52Demo: Memory recall
- 22:12Physical form with Lego
- 22:55Web Bluetooth communication
- 24:42Demo: AI light control
- 26:05Building Arduino car body
- 26:49Live demo: Remote car
- 27:32Integrating car control
- 29:18Complex reasoning demo
- 29:40Summary and demo
- 30:01Why build K.I.T.T.
- 31:26Browser APIs and AI tools
- 32:09Closing and Q&A
related talks
43mLet’s build K.I.T.T. with JavaScript
43mLet’s build K.I.T.T. with JavaScript
30mLet’s build K.I.T.T. with JavaScript
20mFrom ML to LLM: On-Device AI in the Browser
30mLook ma, no hands! Multimodal AI-Agents in the browser
46mFrom ML to LLM: on-device AI in the browser
37mJavaScript On Microcontrollers
31mFrom ML to LLM on device AI in the browser