one-yolo-coreml is a small experiment: take one YOLO detector compiled to CoreML and run it against a wall of dashcam streams at once on Apple Silicon — no per-stream model copy, just one instance time-sliced across every channel. The question was how many feeds it holds before the frame rate falls off.

Each channel draws its own boxes (car, truck, confidence %) and prints its own FPS. Tap a recording to play it in a popup.

1 — Six streams, one model. Every channel holds ~54 FPS.

2 — Same six-up layout on another drive.

3 — Pushed to eight channels: throughput settles around ~44 FPS but every feed stays live.

Takeaway: a single CoreML YOLO instance scales further than expected — six feeds cost almost nothing, and eight still clears real-time. Next I want to see where it breaks, and whether batching the channels into one inference call beats round-robin.