Forewarn

Most flood warnings in Africa cover a whole region, or come too late. Forewarn answers a smaller question: will it flood here, at this pin, this week?

You drop a pin and ask in plain words, or say it out loud in English, French or Swahili. You get a level (green, amber or red), a chance in words like “1 in 5 forecasts”, and what to do.

Code makes every number. It reads 51 rain forecasts for the next 14 days, checks how wet the ground already is, and compares that with the rain that flooded the place before. The model only explains those numbers and picks what to show. A checker compares every number in the answer with the tool results, and if one doesn’t match, the model tries again.

You can follow an area. Each one gets its own agent that checks the forecast every 6 hours and drafts an alert when a flood looks likely. A person reads every alert before it goes out. It’s the same rule I use in my sales agents: the agent drafts, a person approves.

I tested it against 36 past African floods with today’s rules. It flagged 29 of them in advance, 28 of those 3 days or more ahead.

I built it for the Meta Global AI Developer Hackathon, on Meta’s Muse Spark, SAM 3.1 and Muse Voice Transcribe, running on Cloudflare Workers and Durable Objects.