YOLOv9 and Tinygrad turn your CCTV into an AI sentry

What it does

Clearcam is a self-hosted NVR that ingests RTSP camera streams, runs YOLOv9 object detection and tracking locally via Tinygrad, and records event clips. It exposes a browser dashboard at localhost:8080 for live viewing and search, while a companion iOS/Android app relays push notifications and remote clips. An optional premium tier adds end-to-end-encrypted access for use outside your home network.

The interesting bit

The project runs YOLOv9 on Tinygrad—a lightweight deep-learning framework—instead of the heavier runtimes typical in surveillance software. Detection events are also fed through Qwen3 VL to generate human-readable notification summaries, so your phone buzzes with context rather than raw bounding-box metadata.

Key highlights

• Local-first processing: detection and recording stay on your own hardware
• Browser-based event search and playback at localhost:8080
• iOS app ships with zero external dependencies, according to the README
• AI summaries via Qwen3 VL vision-language model
• Optional premium adds encrypted remote streaming and cross-network notifications

Caveats

• Windows support is currently broken pending an ffmpeg fix
• Remote notifications and live streaming outside your LAN require a paid premium subscription
• The README mentions a BEAM=2 performance flag but does not explain what it does

Verdict

A solid choice for developers who want a hackable, fully local AI security stack and do not mind compiling the iOS app from source or paying for remote access. If you need polished Windows support or a completely free cloud relay, look elsewhere.