CivicPulse

CivicPulse turns a neighbour's photo and a few words into a verified civic case that the right team can act on, and it checks the repair before anyone says "fixed". Pilot city: Ambato, Ecuador. The problem: in many cities, reporting a pothole, a broken streetlight or a blocked bus stop means a phone call or a form that disappears. The city gets duplicates, vague text and no proof. When a crew says a problem is fixed, nobody checks. Citizens stop reporting because nothing seems to happen. How it works: 1. Report: no account, no form. A citizen takes a live photo and says what is wrong in their own words, by text or voice, in Spanish or English. 2. Understand: Muse turns the text into a structured spec (category, severity, what it should look like) from a catalog of civic categories. 3. See: SAM 3.1 finds what the spec describes in the photo, and Muse checks each crop. Faces, people and plates are blurred before anything is public or sent to an API. 4. Agree: nearby reports from independent devices become one case through a weighted consensus (device trust, reused photos, impossible GPS jumps), not raw volume. 5. Route and prioritize: the responsible team gets a brief that cites the real reports, with an auditable priority (main road, bus route, school nearby, pothole size measured against the street). 6. Close: the crew or a neighbour sends an AFTER photo. SAM 3.1 and Muse compare it with the BEFORE photo. Doubt becomes "cannot verify", never "resolved". The AI only proposes; people confirm. Results on 63 real, hand-labeled pothole repairs from a public entity in Ambato, anonymized: - 0 of 130 unfixed or mismatched cases marked "resolved". - 0 of 63 repairs from another place accepted, even when the GPS said "near". - 0 of 62 repaired streets mistaken for a pothole when reported. - Unsure cases go to a human instead of being guessed. Judges can run the whole cycle without an account at civicpulse-meta.vercel.app/sandbox, with real before/after pairs.