# Relight

- **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:** [Malitha Bandara](https://cerebralvalley.ai/u/malitha)
- **GitHub:** https://github.com/MalithaBandara/relight
- **Demo video:** https://youtu.be/Wdf_SaKpP8I
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2449

When a storm cuts power, a utility without smart meters learns where the fault is mostly from phone calls, in several languages, while crews wait for a location. Oak Ridge National Laboratory estimates that major power outages cost US homes and businesses more than $67 billion a year, and the average US customer spent 11 hours without power in 2024. Relight is an AI outage agent that turns that flood of reports into located faults, safe switching plans, crew dispatch and customer updates. It works with or without smart meters, so a utility can start from the phone lines it already has.

It runs the whole loop itself: it answers each call in the caller's language (English, Spanish, Hindi, etc.), traces the fault, asks a neighbor when two explanations compete, sends the right crew, revises its diagnosis from crew reports, finds hidden faults after a repair and keeps customers updated. It also reads text messages, damage photos, drone clips and crew voice reports. Calls about a live wire, a fire or a critical customer are escalated, and in live mode go to a human call-taker, whose copilot fills in the report and suggests the next line.

Meta's models read; deterministic code decides. A probabilistic tracer searches every protective device, transformer fuse and line for what best explains who is dark and who is not, using calls, neighbors' silence, smart meters, SCADA alarms and located damage. It opens an incident at 80% confidence and shows the evidence behind it. Restoration switching is simulated step by step and blocked if it would energize a known fault, an unconfirmed fault zone or a crew's work zone, and an operator approves every plan. A hospital whose backup generator fails moves above a live wire in the queue.

Results on simulated storms: Galle found 6 of 6 faults, Houston with failing smart meters 7 of 7, and Chennai with about 2,400 messages across 24,559 customers 9 of 9, with nobody left dark. On an independent IEEE 123-node test with no tuning, it found 4 of 5, leaving only 16 of 1,658 customers dark. Its priority schedule cut customer-minutes interrupted by 30% in Galle and 22% in Houston against first-call-first-served, and in Galle a live wire is made safe after 2 minutes instead of 90. Removing the neighbor question loses 2 of 6 faults in Galle and 1 of 7 in Houston, so the agent's own questions matter.

It is also cheap to run. Model calls cost $3.06 for Chennai, the largest storm, at Meta's published prices. For a 100,000-customer utility we estimate 5 to 9 cents per agent-answered call, phone line included, against a cross-industry figure of about $8 for a live contact (Gartner, 2019).

All code was written during the event. The repository is open source under Apache 2.0, and the live model responses are committed, so anyone can replay the runs without an API key.

## More from Meta Global AI Developer Hackathon

- [Merlin](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2446)
- [La Asistenta](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2447)
- [navfail](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2448)
- [One Market](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2450)
- [IntentLens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2451)
- [Markwise](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2452)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2449. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt