In May I was at the 418 Hackathon at Mahindra University, run by AEON, with Mohammed Abdullah Shareef, Hussain Ali Khan and Mohammed Kaif Pasha. One track, one problem, a ₹1,75,000 prize pool.
The problem we picked: after a flood, victims wait 24 to 48 hours for insurers to process satellite imagery by hand before any money moves. We thought that was too long.
OrbitClaim Sentinel
It takes raw Sentinel-1 SAR imagery, which sees through cloud, and produces an automated flood-claim decision in under five minutes: flooded area in km², a land-use breakdown and a verdict, all in a JSON file under 10 KB. The size is the point. A satellite could downlink the decision instead of 52 MB of raw tiles, a bandwidth saving above 99.9%.
The flood segmentation runs on IBM's TerraMind. On our evaluation it reached an F1 of 0.628 against a baseline of 0.455. That is a real improvement, and still far from something an insurer should pay out on without a person checking it. That is the honest status of most hackathon models, ours included.
The event itself deserves a mention: proper accommodation, good food, and games between debugging sessions. It made the all-nighter easier.