Panopticon Live

Panopticon Live extracts seven biomechanical fatigue signals — recovery latency, serve-toss variance, ritual entropy, crouch-depth degradation, baseline retreat, lateral work rate, split-step latency — from standard 2D broadcast tennis video. No hardware sensors. No depth cameras. No motion capture suits. Just YOLO11m-Pose, a Kalman filter on court-meter geometry, and a 3-pass DAG (forward + RTS smoother + semantic state machine) running on an Apple M4 Pro. The dashboard is a 2K-Sports-style HUD that visualizes the data live — but the FINAL PRODUCT is the downloadable .csv biometric file plus the qualitative match transcripts produced by a three-agent Claude Opus 4.7 swarm (Analytics Specialist → Technical Biomechanics Coach → Tactical Strategist). Tab 3 replays the captured trace with timed pacing, showing the model think, reject hypotheses, and synthesize tactical exploits in real time. Built solo in five days with Claude Code. Twelve project skill packs (cv-pipeline-engineering, physical-kalman-tracking, react-30fps-canvas-architecture, etc.) compose the architecture. MIT license. Live at panopticon-live.vercel.app. The signal is in the infinite pool of pixels on the open internet. Nobody is reading it. We are.