# MIDUSAFE Sentinel

- **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:** [Levis Rabah](https://cerebralvalley.ai/u/midusa)
- **GitHub:** https://github.com/levisrabah/midusafe
- **Demo video:** https://youtu.be/jPTWLEuq3po
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4894

MIDUSAFE Sentinel turns fragmented community emergency reports into structured, traceable disaster intelligence for human responders.

The problem: during floods in Nairobi's informal settlements, reports arrive as scattered WhatsApp-style messages and voice notes in Swahili, Sheng and English. They are duplicated, contradictory and vague about location. Responders cannot tell what is real, and an AI that guesses could do harm.

What Sentinel does: a report (text or voice) goes to the backend, where a Meta Muse model extracts a strict, Pydantic-validated incident record: type, language, translated summary, severity, urgency, confidence, missing information and follow-up questions. Coordinates are never produced by the model. They are resolved server-side from a gazetteer, or the location is flagged "unresolved" rather than guessed. Sentinel then finds possible duplicate reports and explains WHY they may be related (it never auto-merges), detects contradictions between reports (for example one report says a footbridge is impassable and another says it is passable) and sends a human to verify on the ground, and generates an evidence-backed brief in which every claim cites source report IDs (unknown citations and medical advice are rejected server-side). Follow-up questions are drafted in the reporter's own language and placed in an approval queue; a human approves, edits or discards them, and nothing is ever sent or dispatched automatically. Every responder action is audit-logged, and each incident has a Meta AI trace panel showing the model, measured latency and validation result for each step.

Design principle: AI proposes, humans decide.

All demo data is fictional and labelled SIMULATED in the UI and data. Honest measured results on a small 18-report synthetic evaluation: structured-output validity 1.0, incident type 1.0, location resolution 1.0, language detection 0.944, severity-band 0.778, urgency mean error 0.67. This is a small synthetic set, not a benchmark.

Pre-existing vs new: MIDUSAFE existed before the hackathon (basic dashboard, Gemini triage, Leaflet map). Everything Sentinel (Meta integration, extraction, storage, duplicates, contradictions, briefs, verification, approval queue, trace, voice, tests, eval) was built during the hackathon. See HACKATHON_BUILD_LOG.md and the pre-hackathon git tag.

## More from Meta Global AI Developer Hackathon

- [One Cube Up](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4891)
- [Argus Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4892)
- [Open Intelligence](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4893)
- [Zān](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4895)
- [Mathew Ukoh](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4896)
- [HearthAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4897)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4894. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt