Romus

Romus uses MediaPipe to extract 33-point pose landmarks at ~30 FPS from live video, then runs a deterministic biomechanics rules engine to flag form breakdowns. On top of that I built a multi-loop agentic system on Claude Sonnet — with RAG over a curated knowledge base and per-user memory — that generates personalized voice cues during a set and full post-set reports. The backend is FastAPI streaming over WebSockets for low latency, with the Backboard SDK wiring it together.