Intro

From raw sports video to reviewable evidence.

RefereAI MLX turns ordinary match footage into local, inspectable sports intelligence: sport classification, object overlays, frame logs, summaries, and a human review loop.

Watch demo clips Try it yourself
Problem statement

Most amateur sports video is useful, but not searchable or reviewable.

Parents, coaches, and small venues already capture games on phones. The hard part is turning those clips into structured evidence without uploading private footage to a cloud vision stack or requiring broadcast-grade camera placement.
How we solve it

The Mac becomes the local venue brain.

Video enters from a phone, upload, or stream. YOLO26 MLX runs locally, RefereAI observers add sports context, and the result lands in a review console where mistakes can be corrected.

1Phone / videocamera or clip
→
2MacYOLO26 MLX
→
3Observerssport, tracks, score
→
4Reviewclips + feedback
01

Ingest

Use a phone camera, a saved match clip, or a public broadcast segment.

02

Infer

Run YOLO26 MLX on Apple Silicon with no external inference API.

03

Understand

Layer sport classification, ball/player evidence, timing, and overlay rules.

04

Review

Inspect overlays, mark wrong sport or bad overlay, and preserve demo-worthy runs.

Demo

Public clips across volleyball, tennis, and badminton.

Full-width clips for recording the demo: public broadcast or public sample footage, raw links, overlay links, and summary JSON. No private personal footage is used on the public site.

All public demo assets are 720p where source allows. Volleyball clips are 12 seconds; tennis and badminton use checked-in public sample footage.

Try it yourself

Run the same loop locally.

Clone the repo, bootstrap the MLX model, start the local server, then open the app on your Mac or phone.

git clone https://github.com/jravinder/refereai-mlx
cd refereai-mlx
bash scripts/bootstrap.sh
.venv/bin/refereai serve --port 8765
Run docs GitHub repo Local app: /app
Future state

Where this goes next.

Better identity

Track players, teams, referees, spectators, and off-court context with less flicker.

Smarter score

Use scoreboard OCR and rally state so the score changes only when play actually resolves.

Multi-camera rooms

Blend official corner phones and guest phones into one private venue view.

Family viewing

Share a private link with viewer-specific overlays, commentary, and replay controls.

Reach out

Questions, ideas, or a sport to test?

Open an issue, inspect the code, or contact Ravi. The project is public so the build can be reviewed end to end.

Open an issue About the builder