I built a local AI welfare monitoring system that analyze my cat through 4 RTSP cameras on a Mac. No cloud needed, people get masked.
retained text
I didn't want to pay a monthly subscription so a pet camera could tell me what my cat did all day, or send my home footage to someone else's server for it. I used to do animal behaviour research with AI at a zoo, so I built my own. **What it does:** finds the animal, works out which cat it is, logs where and what (posture, resting, walking, eating and so on) on a floor plan, writes a one-line diary for each clip, and builds day, week and month routine reports. When a day drifts from the cat's own usual, it flags it with the reasons. **How it's hosted:** one Mac. Four RTSP cameras. RF-DETR for detection, DINOv2 features to tell the cats apart, and a small vision-language model fine-tuned locally for the diary. **Privacy:** video, detection, identities and reports never leave the Mac. The diary model runs locally by default. There's an optional cloud model, but it only ever sends cropped cat pictures and refuses any clip with a person in it. People in shot are masked on screen, and a camera can be switched off completely. I'd love blunt feedback. What would you want from something like this? Would you trust an AI record of your pet more if it admits what it didn't see, or does that just look unfinished? A Home Assistant integration, Frigate as the detector, exports?
document fetched 2026-10-24 10:32 · canonical https://reddit.com/r/selfhosted/comments/1xdfygb/i_built_a_local_ai_welfare_monitoring_system_that