# Ledgerer

- **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:** [Alexandr Achapouski](https://cerebralvalley.ai/u/AchapouskiA), [Rahim Askarov](https://cerebralvalley.ai/u/RoroDev), [Elmir Abdullaiev](https://cerebralvalley.ai/u/elmircan)
- **GitHub:** https://github.com/RoroDev2023/ledger
- **Demo video:** https://youtu.be/1Ta_UB19WXs
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/90

LEDGER writes the itemized security deposit statement a New York landlord has to send within 14 days of move-out, and it won't charge for anything it can't prove.

The landlord uploads the floor plan, or we scan the room with an iPhone. Gemini 4 Argon turns the plan into rooms, walls, doors and windows, and we build a 3D model from it. Then the apartment is filmed at move-in and again at move-out. Every frame of both videos is pinned to a wall of that plan, so a different camera angle or a moved chair doesn't look like damage.

The Decisions API does the judging as hundreds of small yes or no questions with a probability each: which wall is this, is there a new mark, is it the same spot as before, did only an object move. A charge needs two independent before and after frame pairs that agree. If the proof isn't there the line is $0 and says why. Moved furniture is never charged. Prices and the send-by date are plain code, not model output.

The result is a draft statement with before and after photos for every charge, red pins on the 3D model, and $0 lines that explain themselves.

For the tenant there is a second side. "Find fixes on Shopify" turns each charge into a repair kit: Hydrogen writes the search, the Shop catalog returns real products, the Decisions API picks one, and an agent opens a visible browser and fills each store's real cart. It never checks out. On our test run it was $54.67 of supplies against an $85.00 deduction.

What we measured today: Argon read our test floor plans with every room and opening found and about 1% size error. On our own before and after videos of a conference room, the run found the new item stuck to the east wall, charged it, and marked the stain on the west wall as already there at move-in, in about two minutes per pair. These are small samples from our own footage, not a benchmark.

Limits: pricing uses placeholder rates, plain walls with nothing on them are hard to tell apart, and the statement is a draft for a person to review, not legal advice.

## More from Gemini 4 NYC Hackathon

- [Seen IT](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/87)
- [Pit Crew](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/88)
- [SmartCart](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/89)
- [3D SHOP](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/91)
- [Shopexa](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/92)
- [Systemind](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/93)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/90. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt
