# SunSeg

- **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:** [Rohit Kottomtharayil](https://cerebralvalley.ai/u/bigroko)
- **GitHub:** https://github.com/rkottomt/sunseg
- **Demo video:** https://www.youtube.com/watch?v=378TkVkxVFM
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1065

Grid operators can't see most rooftop solar. Home systems sit behind the meter, so nobody measures what they produce. SunSeg maps it from public aerial photos with no training data.

SAM 3.1, prompted with two words, "solar panels", segments arrays in 0.3 m USGS imagery of Fresno, California. We scored it against Duke University's hand labels on held-out tiles: 90% precision and 81% recall with a lenient overlap rule, 68% and 75% under the standard IoU >= 0.5 rule from the literature, and a pixel IoU of 0.51 (U-Nets fine-tuned to a new area reached 0.60 to 0.62). It costs about $0.44 per km2 of imagery.

Each array becomes a pvlib system: size from its footprint, orientation from its shape, PVGIS typical-year weather. A 3D map shows all 14,710 mapped arrays in Fresno (78 MW DC, 122.5 GWh a year) hour by hour, including the 43.6 MW drop between 3 and 7 pm on 21 June that the rest of the grid has to cover.

Muse Spark then looks at every SAM detection and judges whether it's really a panel. Keeping only what both models agree on raises precision from 90% to 99%. Muse Spark is also the analyst: it calls pvlib itself to answer what-if questions in the app, and every tool call is shown.

## More from Meta Global AI Developer Hackathon

- [DrakenMario](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1062)
- [Nichita Stane](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1063)
- [MoDesign](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1064)
- [Anytime Lab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1066)
- [Adiyan-Wayfinder](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1067)
- [Amitech](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1068)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1065. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt