Local sports vision · YOLO26 MLX · no cloud inference

The private venue brain for amateur sports.

Most sports AI assumes a clean broadcast feed. RefereAI starts with the messy reality: a parent phone at a YMCA, siblings crossing the frame, people sitting courtside, a scoreboard in the corner, and no guarantee the sport is known.

Watch the demo path View source →
Current demo flow
01Raw phone videolive or prerecorded
02Local YOLO26 MLXMac inference
03Venue observerssport, teams, score probes
04HIL Reviewhuman feedback loop
Problem statement

Amateur sports video is valuable, but it is chaotic.

No official cameras.Parents use phones from bad angles.
No stable scene.Kids, spectators, chairs, and other games share the frame.
No cloud comfort.Families do not want youth sports footage uploaded by default.
No review loop.Bad detections vanish instead of becoming training signal.
Purpose

Turn any phone into a private sports observer.

RefereAI is not trying to be Hawk-Eye on day one. It is a local system that moves from not knowing to knowing enough to explain itself: what sport is visible, where players and objects are, what overlays are useful, what the score might be, and which moments deserve human review.
Architecture

Phones are cameras. The Mac is the brain.

Capture

Phone PWA sends JPEG frames over LAN, Tailscale, or a private tunnel.

Inference

YOLO26 MLX runs on Apple Silicon. No cloud inference and no account requirement.

Observer stack

Tracking, sport inference, role heuristics, team color hints, scoreboard probes, commentary, and replay metadata.

Artifacts

Raw video, overlay video, frame logs, summaries, stills, and feedback are stored locally by session.

HIL Review

Human-in-the-loop console groups duplicate videos and opens every analysis version for feedback.

Viewing

A private family room can expose the stream with viewer-side overlay/commentary toggles.

What we built

More than native YOLO boxes.

Unknown → known classifierStarts without a selected sport, accumulates evidence, then locks with rationale.
Multi-sport breadthTennis, volleyball, basketball, badminton, table tennis, pickleball, soccer/cricket candidates.
Overlay rendererPlayers, objects, sport labels, score probes, commentary/call surfaces, and compact HUD variants.
Video workbenchRaw uploads, full-clip analysis, overlay playback, evolution wall, and HIL Review console.
Local privacy modelRuns from Mac, serves local debug/review pages, keeps match artifacts under local folders.
Future steps

The next version gets sharper, not just wider.

A

SAM-style selection

Let parents select one child/player and keep that identity through clips, not just detect generic people.

B

Scoreboard OCR hardening

Separate score candidates from real score changes; rallies should not mutate score until a point event is inferred.

C

Multi-camera venue view

Combine official corner phones and optional parent guest cameras into a better room model.

D

Private cloud relay

Use Cloudflare/Vercel surfaces for sharing links while keeping inference and source capture local-first.