# GitHappens

- **Event:** [Gemini 4 NYC Hackathon](https://cerebralvalley.ai/e/gemini-4-nyc)
- **When:** Sat, Nov 21 at 9:00 AM – 10:00 PM (EST)
- **Where:** New York, NY
- **Team:** [Arth Agrawal](https://cerebralvalley.ai/u/arth21), [Adya Goenka](https://cerebralvalley.ai/u/AdyaGoenka), [Ishita Jain](https://cerebralvalley.ai/u/ishitajain14)
- **GitHub:** https://github.com/ArthAgrawal/Walkthrough_Google_Deepmind_Hackathon
- **Demo video:** https://drive.google.com/file/d/1u_M64Kne1wE9y303wtl-tzPuGKq6nQDy/view?usp=sharing
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/105

Walkthrough: 

A New York City tenant films their apartment and says who lives there, in any language. Walkthrough returns which problems are likely violations of NYC housing rules and which aren't, with timestamped evidence, a plain-language explanation in the tenant's language, and an English complaint draft the tenant reviews and files through 311 themselves.

The problem: tenants often don't know which problems the law covers (a window guard is required only if a child 10 or younger lives there, in a building with 3+ apartments), and a 311 complaint without evidence is the tenant's word against the landlord's.

The answer depends on both the video and what the tenant says. The same unguarded window is a violation with "my son is 6" and not with "no kids live here"; "Run again without audio" shows both results side by side. When a tenant reports a problem, the app finds where that object appears, cuts a 9-second clip with sound, and checks whether the video supports it. If a tenant says "the toilet doesn't flush" but the video shows it flushing normally, the app says so.

Every report is sorted into: you reported it and it's likely a violation; you reported it but it's likely not; we found it on your video; or please confirm / film again. Each card cites the actual HPD order and law (for example, HPD order 790, §27-2043.1), from a reference of 26 conditions built from HPD's own violations data dictionary. A Managed Agent caseworker looks up the building's open HPD violations in the background.

Evidence so far: on real HPD records, Decisions API category answers at ≥ 0.9 confidence were right 98% of the time (41 of 88), and the pipeline routes on that confidence. On a real 70-second walkthrough (4 adults, no children), the 3 unguarded windows are correctly not violations; with audio ignored, the same frames give 1 violation and 2 "film again". A regression check on 4 development videos met all expected outcomes. We don't report video accuracy numbers yet, because we have no labeled held-out video set, and we'd rather say that than report numbers from labels written after seeing the output.

Limits: not all housing law, not legal advice, no health claims, never files for the tenant, and a video can only show what was filmed.

## More from Gemini 4 NYC Hackathon

- [MAB AI Strategies](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/102)
- [Sharms4NYC](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/103)
- [Capture Studio](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/104)
- [StageState](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/106)
- [MoveIn - Columbia](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/107)
- [CodeWarrior](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/108)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/105. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt
