Vellam Kaappaan

Vellam Map is a flood early-warning system for Chennai, built from the photos people already share. During a flood the question that matters is "how deep is this street right now?", but a photo is not data: a driver or a rescue team cannot tell what depth it shows. A user uploads a photo. The app (1) segments the water and every vehicle wheel with SAM 3.1, (2) fits a circle to each visible wheel rim and converts the waterline into centimetres with an error range, using the wheel as a ruler, (3) has Muse Spark independently judge whether it is a real flood and estimate depth by eye, and (4) compares the two. If they agree, the report is published to a live map; if they disagree, it goes to review instead of the map, because a wrong depth is worse than none. Duplicate, stale and wrongly located photos are caught, and faces and number plates are blurred. Reports cluster into area alerts, and people can ask questions by voice in Tamil or English (Muse Voice Transcribe + Muse Spark), answered only from current reports. We tested the real pipeline on 100 public photos of flooded cars with known depth levels. Where Muse Spark gave a depth (69 photos), its rank correlation with the true level is 0.89 and 68 of 69 were within one level. Wheel geometry read only 12 of 100 photos, so it is one input, not the answer. Testing also found that our first flood check wrongly rejected 33% of real floods (23 of 70 held-out photos); we changed it to grade its evidence, which cut that to 1 of 70, at the cost of about a third of real floods showing as "unconfirmed". The dataset gives ordered levels, not centimetres. Depths are estimates with an error range, never a statement that a road is safe. Anyone can try it without an API key: start the backend with REPLAY=1 and each sample photo plays back the pipeline run recorded on it, and the Benchmark tab replays the 100-photo run, intermediate images included. Stack: FastAPI, SQLite, React, Leaflet. Built during the event.