Leila Vi

FireRisk Intelligence: finding the fire risks hidden between a building’s everyday alerts In a real building, fire risk rarely shows up as one big alarm. It builds up across records that nobody reads together: an electricity spike, an overdue panel repair, a missed generator run, a blocked extinguisher seen on a walk-through video. Each looks minor on its own. Together they can mean the building would not stay safe during a power cut. I manage operations for a 6-floor commercial building in Vietnam. FireRisk Intelligence reads our real operational records, with written approval from the data owner: meter readings, incident and repair logs, generator and diesel logs, asset registers, checklists and on-site survey videos. It turns them into one dashboard that explains which fire-safety risks are building up, where they are, and why. How it works: Deterministic rules decide the risk level. Meter spike and drop detection, nine fire-hazard pathway rules, and a compound-risk rule set every alert level. AI never changes a level. Meta Muse Spark 1.3 investigates each alert as an agent. It uses tool calling over 7 read-only tools to look up records, then returns a strict JSON briefing: situation, competing scenarios with supporting and weakening evidence, limitations and actions. Every citation is checked against real record IDs, and claims without a valid record are downgraded. Muse Spark reviews the survey videos. It has reviewed 14 real videos and produced 34 timestamped fire-safety findings. SAM 3.1 visually grounds those findings in video frames, so a reviewer can see where the problem is. The app also reports honestly when SAM finds nothing. A “butterfly effect” chain links the risks together. For example, the building has one shared ATS, a generator that powers all 6 floors, and emergency lights with batteries of uncertain age. The chain shows how a small equipment fault becomes a life-safety risk. Privacy by design: Personal data is pseudonymised before any AI call. Technical drawings never leave the local machine. Faces and signs are blurred in demo frames. The public repo contains code only. A fictional sample building (“Acme Riverside Plaza”) lets judges run the full dashboard. AI analysis runs only with the judge’s own Meta API key and is never faked. Track fit: Build for Your World. All code was written during the hackathon. My earlier prompt-only workflow, which contained no code, is disclosed in the repo.