What it does Sentinel is a three-tier system: an Android app (CamFlow) streams JPEG frames over your LAN to a Flask server, which serves an MJPEG preview in the browser, records segmented MP4s, and optionally triggers AI analysis. No cloud required, no SD cards, no subscription — just a Wi-Fi network and a phone you already own.

The interesting bit The architecture treats the phone as a dumb frame pump and keeps all intelligence on the PC. A shared FrameBuffer decouples streaming, recording, and AI so none block the others. The AI path is deliberately two-stage: traditional CV (motion detection) acts as a cheap gatekeeper before calling in a vision model for structured semantic output — person count, activity, risk level, confidence — all logged as JSON events.

Key highlights

• Android client auto-discovers the server via UDP, so you don’t hard-code IP addresses
• Browser dashboard controls preview, recording, and AI parameters without installing anything else
• Pluggable model interface: swap in your own vision backend via config
• Segmented local recording with time-based file rotation for long-term runs
• Structured JSON event logs for downstream analysis or model fine-tuning

Caveats

• Multimodal AI inference is explicitly recommended to use an online service, so “local-only” applies to the pipeline, not necessarily the cognition
• The README notes this is positioned as both a tool and an “engineering prototype platform” — expect to tinker
• Android 8.0+ required; no iOS path mentioned

Verdict Grab this if you have a drawer of old phones and want a hackable, privacy-respecting vision pipeline for home labs or small-scale data collection. Skip it if you need turnkey cloud analytics or a polished consumer app.