One Fusion

StillWater turns a phone video or photo into a mosquito-breeding-site inspection. A volunteer films a rooftop or street; StillWater finds every container holding water, ranks each by risk, and writes cleanup work orders in Bangla and English with photo evidence. Problem: Dengue mosquitoes breed in water held by tyres, buckets, pots and drums. Removing these containers is one of the most effective ways to stop the spread, but inspectors still check by eye, one roof at a time. In Dhaka, where I live, this happens every monsoon. How it works (7 steps): 1. Ingest the video or photo 2. Privacy: SAM 3.1 finds people and blurs them before anything is saved 3. Scout: Muse Spark Open looks at the scene and picks which objects to search for 4. Track: SAM 3.1 segments and tracks each container across the video, so each is counted once 5. Verify: two different Meta models judge each container independently; if they disagree, it is flagged for a person to check 6. Decide: fixed, published rules assign the risk level, so the AI describes and the code decides 7. Act: an agent writes bilingual work orders and plans a rescan Meta models used (100% open weights): SAM 3.1 [where it runs], Muse Spark Open on Baseten, [second model + where it runs]. No training data, no drone, no paid AI account needed, so any city or NGO can run it. Honest limits: small, dark or hidden containers are missed more often, and seeing water inside a container is hard, so we say "risky container", not "confirmed breeding site". Results are AI estimates to be checked on site; this is an environmental cleanup tool, not medical advice. All code was built during the hackathon (Nov 2–4, 2026). Live demo: https://stillwater.onefusionlab.systems/ Video: https://youtu.be/D3YvXPksqRU Code: https://github.com/AdribMahmud101/stillwater