How the AI sees your shot,
frame by frame.
Coach AI combines YOLO computer vision with a PyTorch Transformer trained on cricket pose data — running on Azure, delivered through a React app you open in any browser.
Clip upload to coaching report in four stages.
YOLO-Pose & object detection.
Each video frame runs through Ultralytics YOLO-Pose to extract body keypoints, plus custom detectors that find the striker, bat, and ball — even in busy nets footage.
Those signals power both movement coaching (head stillness, footwork, balance) and stroke detection when the ball meets the bat.
From keypoints to shot motion.
Instead of raw pixels, we train on pose trajectories — a compact, angle-robust representation built from each stroke window. Left-handers are mirrored to a canonical stance automatically.
Stroke window
We isolate ~2 seconds around bat–ball contact — pre-load, impact, and follow-through.
Multi-ball sessions
For nets videos, motion peaks and contact scores find each ball; every stroke gets its own window.
Annotated replay
OpenCV and ffmpeg render an H.264 MP4 with skeleton overlay you can watch in the browser.
PyTorch Transformer classifier.
A lightweight Transformer reads the pose sequence and predicts the shot type — cover drive, square cut, pull, sweep, defensive, and more. Trained on professionally annotated cricket pose footage.
Secure ML on Azure, results in your browser.
Upload to API
Your clip is sent over HTTPS to our Flask API on Azure App Service. Models run inside Docker — no installs on your device.
Profile & history
Sign in with Supabase to save analyses, track trends, export PDFs, and compare good vs mistimed clips over time.
Always improving
We retrain on cleaner labelled data and deploy new weights without you changing apps — the same pipeline our ML team uses in research.
What Coach AI is built on.
Want to license this stack?
We license our cricket coaching pipeline for academies and sports-tech products. Talk to us about your use case.