# ninemomo - **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon) - **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST) - **Where:** Online - **Team:** [hyeonmo Gu](https://cerebralvalley.ai/u/ninemomo) - **GitHub:** https://github.com/Hyeon-moGu/CiviFold - **Demo video:** https://youtu.be/wcGPEWtgwMY - **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery - **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2347 CiviFold helps city 311 staff tell a duplicate report from a problem that is new or has come back. The problem. Duplicate reports create repeated review work for 311 centers. In September 2026, NYC 311 closed 454 of 3,662 street pothole requests as duplicates of an earlier one (NYC Open Data). A location-only matching rule can confuse a similar fixture nearby with the original, or mistake a problem that returned after a repair or cleanup for one already resolved. That risks overlooking a problem that still needs attention. What it does. A resident reports with a photo, a message or a voice note. CiviFold reads the words first: possible emergencies go straight to the top of the staff queue, and complaints about people or behavior are never matched by image. Reports without photos are not automatically matched. For a photo, CiviFold searches for open and recently resolved incidents within up to 40 m. SAM 3.1 segments the fixture or damage; for road damage and dumped piles, Muse Spark first identifies the region for SAM. Spark compares whole photos and object-and-surroundings crops without seeing timestamps. Rules then apply the observation time and repair history to suggest: duplicate, late report, already fixed, possible recurrence, new problem, one question to the resident, or staff review. Nothing is merged or closed automatically. Staff see the photos, segmentation outlines and comparison evidence, then approve or choose another action. On request, an experimental Spark agent looks for evidence against a closing suggestion and can only make it more cautious. Another writes a case brief with references to recorded evidence. Results (suggestions, not staff decisions). On its first run on a pre-registered holdout of 80 synthetic reports across 16 previously unseen assets, CiviFold suggested linking 34 of 39 repeat reports to the correct incident, versus 30 for our location-only baseline. It made no wrong merge or closure suggestion on the 33 new or returning problems, versus 24 for the baseline. Fourteen correct links required an answer from a simulated resident; earlier test sets had 1 to 3 wrong merges or closures each. These first-run figures precede a later update to road-damage and trash localization. In small, hand-screened checks using real San Francisco 311 photos, CiviFold suggested closing none of 13 requests made after a cleanup. Among 33 pairs judged to show the same problem, it suggested merging 14, asked or deferred 17, and incorrectly suggested 2 as new problems. Our baseline does not represent a city's actual system. No city integration or study of staff time savings has been completed. The live demo includes real San Francisco 311 cases, generated hard cases, a resident form and the staff console. The application code, prompts, synthetic dataset and evaluation harness were created during the hackathon. Real photos come from public San Francisco 311 records. ## More from Meta Global AI Developer Hackathon - [Foal](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2344) - [Tiwa](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2345) - [Resonance](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2346) - [NodXAI Science](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2348) - [Franck Melka](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2349) - [SoloTraveller](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2350) --- Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2347. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt