# relief-net

- **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:** [Francisco Javier Saeta Hidalgo](https://cerebralvalley.ai/u/Saeta)
- **GitHub:** https://github.com/SapSaeta/relief-net
- **Demo video:** https://youtu.be/9CWJKZulYWw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1111

relief-net is an AI coordination layer for community disaster response. After a flood, fire or storm, requests for help arrive scattered across calls, texts and photos, and volunteer coordinators spend hours reading, deduplicating and prioritizing them by hand. With relief-net, an affected person records a phone walkthrough of their damaged home and narrates what happened. The app turns that video into a prioritized, evidence-backed help request, and a human coordinator approves every dispatch.

How it works: (1) Intake: Muse Voice Transcribe transcribes the narration, SAM 3.1 segments damage in 8 keyframes, and Muse Spark 1.3 watches the frames, proposes extra damage for SAM to localize, and writes a strict-JSON help request. (2) Verification in code: every citation must appear in the transcript (3+ words) or match a SAM detection in that frame. "Verified means grounded, not true." (3) Triage: Muse Spark proposes duplicates and a priority queue with a reason per request; code rules decide automatic merges and safety ordering (112 > critical > high > medium > low). (4) Matching: a tool-calling agent (find_volunteers, check_supplies, propose_assignments, flag_for_coordinator) proposes assignments under hard rules (structural risk needs an engineer, electrical risk needs an electrician, medication transport only logistics, never medical advice). Nothing is dispatched without coordinator approval. (5) Reports for insurance/aid with verified citations; unsupported claims are marked "unverified". An append-only audit log records every decision.

Measured results on a synthetic "Riverside flood" scenario (14 requests, 10 volunteers, 8 supply types): both Muse Spark 1.3 (Meta Model API) and Muse Spark Open (Baseten) scored 7/7 checks in 5 of 5 runs, finding 2/2 duplicates with 0 false merges (44 s and 23 s), versus a human doing it once: 2/7, 0/2 duplicates, 3 false merges, 2.8 min. A spoken prompt-injection test never moved a request above a more urgent one (5/5 per provider). One intake costs about $0.17 and ~47 s. All data is synthetic; this is a personal project.

## More from Meta Global AI Developer Hackathon

- [ADADIVE](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1108)
- [PrinceDevs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1109)
- [Zevora](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1110)
- [TestForge AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1112)
- [OrbitRoute](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1113)
- [ChoiceGap](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1114)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1111. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt