Building Coach AI: YOLO-Pose and Transformers on Azure

When we set out to build Gaara AI, the goal was simple: deliver coach-grade biomechanical feedback from an ordinary phone clip — no lab, no installs. Today Coach AI runs YOLO-Pose for body tracking, custom YOLO detectors for player/bat/ball, and a PyTorch Transformer trained on cricket pose sequences.

Each stroke is converted into a keypoint trajectory around bat–ball contact. The Transformer predicts the shot type across 15+ classes — cover drive, cut, pull, sweep, and more — while separate biomechanical rules score head position, footwork, and balance.

Video is processed on Azure App Service inside Docker. OpenCV and ffmpeg render annotated H.264 replays you can watch in the browser. Sign in with Supabase to save history, compare good vs mistimed clips, and track progress over time.

Privacy matters: we process your upload for analysis and return results — we do not sell raw video. Optional training-data collection is opt-in only.

Our biggest lesson for sports ML: invest in detection and labelling before chasing bigger models. Clean pose windows beat raw RGB every time for sample-efficient shot recognition.