Built by Rithvik Burki and Arav Dharnikota at the VoiceOS Hackathon, August 2026.
Say "Hey VoiceOS, open a whiteboard and explain how transformers work."
A glass lightboard slides open beside your call, and a live AI tutor starts drawing while it talks. Diagrams build stroke by stroke on the glass, narrating in realtime, stopping when you interrupt, answers, and adapts the lesson to what you already know. It feels less like an app and more like a teacher standing behind a pane of glass.
- Summoned by voice. No app-switching, no setup, one spoken sentence births the board next to whatever you're doing.
- The AI actually uses the whiteboard. This isn't a chatbot with a canvas bolted on: the model drives the board through live tool calls (draw, write, highlight, clear scene) while speaking, so the drawing and the explanation are one continuous performance.
- It remembers you. Every stroke and every lesson lives in a reactive backend, the board state syncs live, and the tutor's adaptive memory means the next session starts from what you already covered.
- Glass-lightboard design. Modeled on how the best human explainers teach on video: the tutor "writes on glass" facing you, not on a slide deck.
"open a whiteboard…" voice command, VoiceOS deeplink
│
▼
Electron glass lightboard the native-feeling Mac canvas (app/)
│
▼
OpenAI Realtime voice loop the tutor talks, listens, interrupts gracefully
│ live tool calls draw / write / highlight / new scene
▼
Convex reactive backend live board state + adaptive lesson memory (convex/)
- app/, the Electron glass-lightboard frontend: renderer, scenes, stroke engine.
- convex/, the Convex backend: schema + functions, Realtime token action, prompts, VoiceOS launch.
- docs/, the build specs, start with
NORTH-STAR.md·MAIN-STATEMENT.md·SCOPE-LOCK.md. - report/, deliverables and studies.
Clone, npm install in app/, add your OpenAI Realtime + Convex keys, and launch, or watch the live demo, sped up and you'll understand the whole product in a minute.