# GridSense

- **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:** [Abubakar Yahya Hamza](https://cerebralvalley.ai/u/AbuMaheer)
- **GitHub:** https://github.com/AYHHamsuky/gridsense
- **Demo video:** https://youtu.be/GS5NoWbw2vA
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1930

GridSense is a Hausa/English AI agent that turns floods of electricity outage reports into verified repairs, and tells every customer when their power is back.

Live demo: https://gridsense.187.124.112.86.sslip.io/

The problem: in northern Nigeria, one burnt transformer can trigger hundreds of calls and messages, many in Hausa. Distribution companies log the same fault again and again, technicians are dispatched late, repairs are closed without proof, and customers rarely hear back. I work in ICT at a Nigerian electricity distribution company and see this every week.

What it does:
1. Customers report by text or voice note, in Hausa or English.
2. A Muse Spark agent with tool calling works out the fault type, finds the feeder, and either opens a new fault or merges the report into an existing one.
3. It dispatches a technician with the right skill, preferring one who covers that feeder, and replies to the customer in their language with a fault ID and safety advice.
4. In outage spikes (storm mode), a hybrid engine sends routine reports through Muse Spark Open (open weights, served on Modal) in batches of 25; low-confidence, unlocatable or new dangerous reports (fire, fallen wire) escalate to the full Muse Spark agent. A burst of ~150 reports collapses into a handful of faults.
5. The technician uploads a repair photo. SAM 3.1 locates the transformer, pole or meter, and Muse Spark judges whether the repair is complete and safe. Only then does the fault close.
6. Every customer on that fault gets a restoration message in their own language; an unsafe repair triggers a follow-up warning instead.
7. Every tool call and verdict is logged so supervisors can audit the agent.

Built with FastAPI + Streamlit during the hackathon (Oct 27 - Nov 4); the commit history shows the timeline. All data is synthetic and demo photos are AI-generated; no real utility or customer data is used.

## More from Meta Global AI Developer Hackathon

- [WeISM](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1927)
- [SH-01](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1928)
- [AccessFix AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1929)
- [John Ayandipe](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1931)
- [TheFuture](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1932)
- [DevConquerors](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1933)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1930. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt