# Find My Dog

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Aman Sah](https://cerebralvalley.ai/u/Kris_sah)
- **GitHub:** https://github.com/shahaman098/find-my-dog
- **Demo video:** https://www.youtube.com/watch?v=HVPD8go8QDU
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3633

Your dog slips out of the gate at 6:15pm. By 6:40pm, three doorbell cameras on your street have already seen her, and nobody knows.

Find My Dog turns that footage into a search. You file a report with a photo of your pet and a quick voice note. Neighbours upload clips from whatever camera they have: a Ring export, Nest, CCTV, a dashcam or a phone. SAM 3.1 tracks every animal in every clip and keeps each one's ID, so a dog crossing a driveway becomes one candidate instead of 300 frames. Muse Spark compares each candidate with your photo and says why it does or doesn't match. You confirm the real sightings, and you get a timeline of where your pet was seen, a plan for where to search next, and a poster you can print or share.

Why it matters: in the ASPCA's national survey, 93% of lost dogs made it home, and searching the neighbourhood was the most common way owners found them. That search is still done on foot. Ring's Search Party showed people want camera help (Amazon says it brought 99 dogs home in its first 90 days), but it only scans Ring cameras, only looks for dogs, and is on by default. Find My Dog takes footage from any camera, also searches for cats and other pets, and only runs when an owner asks.

Does it work? It's live, and there's a demo report anyone can open: https://find-my-dog.kiritokris.workers.dev/#/demo
I tested it on real footage:
- Against look-alike dogs, the right dog ranked first 9 times out of 9, and 92% of 87 match decisions were correct.
- With clips degraded to look like doorbell footage (night-vision IR, quarter resolution, motion blur, dusk), the right dog still ranked first in 18 of 18 cases.
- SAM kept the same ID on a dog through a 1.5-second full blackout.
- A person SAM had labelled as a dog was rejected 9 times out of 9.
- All 21 match verdicts I ran through the real interface were correct.
Precision on its own is 83%, which is why the app never confirms a sighting without the owner. The test set is small (3 dogs), and I say so in the repo.

Privacy is enforced in code: there's no person or face recognition anywhere, only the owner can run matching or confirm a sighting, neighbours can only add clips, and a withdrawn clip is deleted from storage.

The raw data, screenshots and logs behind every number above are in EVIDENCE.md in the repo, along with what failed and the bugs my own benchmark caught.

## More from Meta Global AI Developer Hackathon

- [Manifest0 AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3630)
- [Omata](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3631)
- [Haven](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3632)
- [MuseGuard](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3634)
- [Saath (Sticking Together)](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3635)
- [The Virtual Neuro Lab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3636)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3633. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt