# FrostWatch Altiplano

- **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:** [Marvin Mollo](https://cerebralvalley.ai/u/magvin24)
- **GitHub:** https://github.com/MarvinMR24/altiplano-frost-alert
- **Demo video:** https://www.youtube.com/watch?v=xzRdJmWZpRw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3394

FrostWatch Altiplano is a mobile app that warns farmers and families in the Bolivian Altiplano (La Paz to Oruro corridor, around 4,000 m) when radiation frost is likely tonight in the exact places they care about.

On clear, calm nights the ground loses heat fast and temperatures drop below 0 °C. One frost can burn potato and quinoa leaves, stress llamas and alpacas, freeze water pipes and ice roads. Most warnings are generic and rarely speak about a specific field or village.

How it works:
- Users find their community on satellite imagery, draw their own zones (houses, fields, corrals) and choose how early to be warned (30, 50 or 70% risk).
- A reproducible ML pipeline uses 41 years of Copernicus ERA5-Land reanalysis (1985 to 2026) for 8 points in the corridor. A transparent logistic regression predicts frost for the coming night from evening conditions up to 18:00.
- Strict chronological split by whole years (train 1985-2009, validation 2010-2017, test 2018-2025). The test set was evaluated once, at the end. On those unseen years it caught 96% of frost nights (recall 0.964) with PR-AUC 0.893 and F2 0.880, above the best simple baseline (climatology, PR-AUC 0.809). Honest trade-off: precision 0.65, so some false alarms.
- When a night is processed, the backend creates alerts per zone and sends them by push and email. Each alert includes a plain-language explanation written by Muse Spark (why tonight is risky, what to do for crops, animals and water), with every number checked against the model output.
- The app works in Spanish, English, Aymara and Quechua (partial translations, under native review).

Honesty notes: the data is reanalysis, not raw satellite imagery or a live forecast, and labels come from the same reanalysis. The demo replays a real night from the test period (17 Aug 2024). Built with FastAPI, SQLite, Docker, Expo (React Native), DVC and MLflow on DagsHub.

## More from Meta Global AI Developer Hackathon

- [AMCA Collective Intelligence Lab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3391)
- [Cifko](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3392)
- [Gwenhael](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3393)
- [MolDeNovo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3395)
- [NKNGO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3396)
- [Scene](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3397)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3394. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt