Team MagIA

After a flood, the people deciding where to send help work with scattered, unverified information: Spanish voice notes, loose photos and rumors. Sending boats to a false report costs time real victims don't have. Centinela verifies each community report against evidence, ranks the zones that need help first, and drafts alerts for a person to approve. A voice note is transcribed with Muse Voice Transcribe (biased with local place names); Muse Spark 1.3 orchestrates the pipeline with tool calling, extracts a structured report, reads a bulletin from a PDF, describes photos and drafts the alert in Spanish and English; SAM 3.1 segments what it sees in a photo, and code measures the share of the image. Code decides every verdict (confirmed, contradicted or no evidence) from the evidence, and every claim cites an evidence ID. The model only explains it, and if it disagrees with the code the report goes to human review. Text inside a report or a PDF is treated as data, never as instructions. A photo alone never confirms a report. Nothing is sent without a person approving it. Honest status: the scenario, evidence, bulletin and demo photo are simulated, while the model calls are real. The 15-case evaluation (0 false confirmations) is a regression check I wrote myself, not an independent benchmark, and I have not yet validated the problem with civil-defense practitioners. Built solo from Pucallpa, Peru, between 1 and 4 November 2026. Open source (MIT).