Back to AI products
Open explain.afdev.site
Next productTounsi
AI2D
Multi-agent AI whiteboard
Ask a question out loud: a multi-agent system draws animated diagrams on an infinite whiteboard and narrates them with a real-time voice, in English or French.

What I built
- Multi-agent system on Claude Code (headless): a drawing agent (Sonnet / Opus), a research agent (web search) and a note-writing agent.
- Streams a custom tag protocol parsed token by token into animated SVG scenes on an infinite 2D plane: first stroke on screen in 2 to 4 s.
- Agent guardrails: per-agent tool sandboxing (file reads confined to the user's folder), generated-SVG sanitisation, SSRF protection, search and credit caps.
- Real-time voice: streaming Gemini TTS with key rotation and local fallback (Piper, Kokoro), Whisper STT, voice barge-in, speed control that keeps the pitch.
- Premium users upload PDFs or links that the agent reads before explaining; every answer becomes an exportable study note.
- Per-question cost tracing in Langfuse (USD 0.035 on average) behind Free / Plus / Premium plans with Stripe and Konnect billing.
Stack
- TypeScript
- Node.js
- WebSocket
- SVG
- Claude Code
- Gemini API
- Python
- faster-whisper
- Langfuse
- Docker
