# Watch Me

- **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:** [Kibaek Kim](https://cerebralvalley.ai/u/hyper07)
- **GitHub:** https://github.com/hyper07/meta-hackathon-2026
- **Demo video:** https://www.youtube.com/watch?v=mupz2IrtyWo
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/241

Try it live, no install or key: https://kibaek-kim2018--watch-me-demo-web.us-east.modal.run

Watch Me is a personal safety assistant for harassment, assault and domestic violence. Say "Watch me" before a party, a date or a walk home. Only your voice, checked against a voiceprint, can turn it on or off or answer "I'm OK"; setup is one line, said once. A team of Muse agents, led by Muse Spark Open, Meta's open-weight model, reads what is said around you. When something sounds wrong (coercion, threats, pressure to drink, signs of drugging), it privately asks "Are you OK?" If you don't answer, say "no" in your voice, press "Get help" or say your SOS phrase, it texts, calls and emails the people you chose, with what it heard and a map link that updates as you move. A stranger or partner who says "stop watching" is refused, and that moment's audio is kept as evidence. A duress phrase makes it look off while it keeps listening and alerts.

Someone who has been drugged can't say "I'm not OK," and at home, reaching for a phone isn't always safe: more than 1 in 3 US women have faced violence or stalking by a partner (CDC), and 44% of domestic violence never reaches the police (BJS). At school or work, harassment often hides inside a joke. Notice mode is for that: tell it what happened, and it shows which exact words match how the EEOC describes harassment, keeps a dated, hash-chained record, and counts patterns per person. It then shows similar real public cases with their official outcome, verified places to turn and your own Title IX or HR contact, and exports the record for you to share. Legal information, never legal advice.

It works on the web and on iPhone and Android with the screen locked, and offline. On 22 realistic test scenes, Muse Spark Open asked in every harmful scene given a perfect transcript, with no needless check-ins, exactly like the hosted Muse models, but reached "Are you OK?" 6 s sooner (9.6 s vs 15.0 s on the live demo). Docker, 76 tests in CI, Apache-2.0.

Next, Meta's wearables: on the Muse Charm pendant it runs in stealth, showing only "Are you OK?" (I built a stand-in on an ESP32-C6 board), and on Meta's AI glasses it works hands-free, with the check-in spoken privately in her ear.

## More from Meta Global AI Developer Hackathon

- [FaultForge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/238)
- [NightMoose (solo: Alex Suarez, GitHub @nightmoose)](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/239)
- [Vestire AI Listing Studio](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/240)
- [VERA](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/242)
- [The Dealer Social](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/243)
- [86TO88](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/244)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/241. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt