Workbench
Labs
Things I build to understand how they work, and then keep running. The code is open to read. The demos are live against real models, so they open with an access code.
2 developments
2026-11
Profe Digital — Voice Tutor with a Live Whiteboard
A fully digital, empathetic teacher for math, physics, chemistry, biology, programming and statistics. You talk, it answers out loud and draws by hand as it explains: step-by-step solutions with a laser pointer, formulas and chemical reactions, code, tables, graphs, photos and simulations. You can interrupt it, it reviews what you write on the board, and it remembers you: it builds your profile by talking, tracks mastery per skill, celebrates every milestone and prepares the workshop you bring, guiding you without giving away the answers.
Study Bidirectional streaming with Gemini Live straight from the browser: a single-use ephemeral token locks the model, prompt and tools server-side, and non-blocking tools update the board while the voice keeps going, with barge-in and proactive audio. The repo also ships a Google ADK 2 backend.
- Gemini Live (3.8)
- Ephemeral tokens
- Excalidraw
- Web Audio Worklets
- Next.js 15
- Google ADK 2.14 (repo)
- HyperFrames
Watch the demo
Demo · needs the access codeSource code (opens in a new tab)2026-11
Eyes & Try-on — Edge Vision Agent
Real-time on-device vision (gaze tracking, hand gestures, selfie segmentation) with a voice agent. Test clothing colors, hair styles, and accessories locally with zero cloud latency, plus Anthropic Claude multimodal streaming.
Study Ultra-low latency edge vision architecture: MediaPipe on-device triage (0 tokens wasted), small JPEG/WebP payloads, and neural voice synthesis.
- MediaPipe Vision
- OpenCV.js
- WebAssembly
- Claude Opus 5.5
- Claude Haiku 5.5
- Web Audio / Neural Speech
- Next.js 15
Demo · needs the access code