Nimbus AI

I was on the bus to college in Tupungato when hail started pouring down. In Mendoza that means harvests ruined in minutes, dented cars and broken roofs. I wondered whether any app warned people in time, and the answer was: barely. After months of reading papers and hunting for data, I found a single major precedent, with a partial dataset. The real problem wasn't the model: nobody had an orderly record of which days hail actually fell.

The project had two stages. The first was exploratory: an analysis and an unsupervised model to pass the first Data Science module at Coderhouse. The second was my end-of-year project in the software degree: a full year and over 300 hours. I built the dataset by pulling weather data from several sources, with scraping and manual checks, and labelled hail days year by year from 2000 to 2024. I added GOES-16 satellite images from the point they became available. With that I trained two networks: a dense one that reads the weather data and a convolutional one (CNN) that reads the images, and their predictions are combined into a single probability. On top of it I built the platform, with a dashboard for each role: civil defence, meteorologists, data scientists, admins and the general public.

As far as I could find, the resulting dataset is the largest labelled hail record in Argentina. On the test data, the model caught every real hail event, at the cost of only one in seven alerts ending in hail. I tuned it that way on purpose: a false alarm costs far less than hail with no warning. Today the platform is live and the model server is paused for cost reasons, waiting for Nimbus 2.0.