# HereO

- **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:** [Jacob Clifton](https://cerebralvalley.ai/u/mememe)
- **GitHub:** https://github.com/jclift3/rork-here-o/tree/meta-hackathon
- **Demo video:** https://youtu.be/3ZCZHsiHSI4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5269

HereO is a community safety network: hold one button and the people you trust, plus trained volunteer heroes within a few minutes' walk, know where you are within seconds. Heroes don't confront anyone. They show up, stay with you, walk you somewhere safe and call 999 or 911. We're launching with students in Cambridge, UK.

For this hackathon we built Muse Dispatcher, an AI response coordinator on Meta's models. A Muse Spark 1.3 agent runs the response after an alert. It reads what's happening, picks up to three nearby heroes, gives each one a single safe job, decides when to push for an emergency call and writes the calm brief the frightened person reads. The person can hold to talk instead of typing, and Muse Voice Transcribe turns panicked speech into text, telling voices apart. They can share a photo of where they are: SAM 3.1 finds every face and number plate and we blur them before anyone sees it, then Muse Spark describes the place ("outside the Bedouin restaurant on Mill Road") so heroes can find them.

Every tool call the agent makes passes a deterministic guard. The agent can add help but never remove it, never lower urgency, and never send a hero to chase or confront anyone. Each verdict (allowed, adjusted or blocked) is logged and shown in the app, so every incident has an audit trail. A 24-scenario eval, including prompt-injection attacks, measures the model's own judgement before the guard steps in.

Live eval at the app's real 10-second budget: 18 and 20 of 24 scenarios passed across two runs, emergency-call recall of 83% and 92% from the model alone, a median agent latency of about 7 seconds, and the guard blocked every unsafe hero job the model attempted ("intervene" in a fight). The first untuned run passed only 3 of 24 because each reasoning step took about 4 seconds; switching to a one-turn plan at minimal reasoning fixed it.

Built during the event (Oct 26 to Nov 4) on the meta-hackathon branch: Muse Dispatcher, voice, scene photo, safety guard, eval and the redesigned live incident screens. The help button and circle alerts existed before.

## More from Meta Global AI Developer Hackathon

- [Heliox](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5266)
- [THE NANOVERSE](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5267)
- [AuraAI-AgenticRealtimeCV](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5268)
- [Digital Dragons Incorporated](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5270)
- [TEREX](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5271)
- [Andrew Rice](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5272)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5269. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt