About

Why this exists.

RefereAI MLX is a local-first sports vision system for the places that do not have broadcast cameras, vendor installs, or permission to upload kids' games to a cloud model.

Purpose

Make messy amateur sports video useful.

For parentsRecord, replay, and privately share a game without handing youth footage to a vendor by default.
For coachesGet a quick second observer for positioning, rallies, score candidates, and review moments.
For buildersShow how YOLO26 MLX can become a local product surface, not just a bounding-box demo.
How it was built

One local loop: capture, infer, observe, review.

Phone capture

Browser/PWA camera sends frames to the Mac over the local network.

MLX inference

YOLO26 runs locally on Apple Silicon through an MLX inference path.

Observer layer

Tracking, sport classification, overlay rendering, score probes, and commentary attach context.

HIL console

Every analyzed clip can be replayed, compared, corrected, and marked demo-worthy.

Credits

Standing on the right shoulders.

  • YOLO26 — Ultralytics shipped the YOLO26 family in January 2026. We use the stock yolo26n weights with zero fine-tuning.
  • YOLO26 MLX baseline — the MLX port provides the local inference foundation. Our tracker layer builds on top of that path.
  • MLX — Apple ml-explore for the array framework, unified-memory inference path, and the Apple-Silicon-native programming model that makes this whole architecture work.
  • ByteTrack — Zhang et al., 2022 for the multi-object tracking algorithm that yolo26mlx/trackers/ implements.
  • Design lineage — warm-paper palette and Fraunces/Inter Tight/JetBrains Mono stack borrowed from the sibling product AISOFT for visual family consistency.
Get involved

Two ways in.

If you want to run it yourself, open the demo page, clone the repo, run bash scripts/bootstrap.sh, then start refereai serve.

If you want to contribute, start with AGENTS.md and INTERFACES.md.

View on GitHub → Read the docs
Built by

Ravi Jilkapally.

RefereAI MLX is built by Ravi Jilkapally, a GenAI product and engineering leader working on practical AI systems beyond the demo stage.

The project grew out of a simple question: can a normal Mac and a normal phone make messy amateur sports video useful without uploading every frame to someone else's cloud?

GitHub LinkedIn AISOFT