# Sheetsight

- **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:** [Philipp Thun-Hohenstein](https://cerebralvalley.ai/u/pthun), [Rahul Modugula](https://cerebralvalley.ai/u/Rahulmodugula)
- **GitHub:** https://github.com/RahulModugula/gemini-hackathon
- **Demo video:** https://drive.google.com/file/d/1EDTo0_teISt5m6eEbUMtQkJV7rQuWXXl/view?usp=sharing
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/80

A single building passes through dozens of hands: the owner, the architect, engineers, the general contractor, and every subcontractor. Over a multi-year project they all work from the same drawings, and the drawings keep changing after they're issued. Every change ripples out through email and PDFs, and every misunderstanding becomes a formal question to the architect, an RFI (Request for Information). RFIs are reported at about $1,080 each (which is fine to an extent) but leads to delays from 8 hours to 7 days each (Navigant, via Procore), and roughly 1 in 10 asks something the documents already answer. In a business that runs on thin margins, that's money lost to paperwork, and worse when someone builds from the wrong version and the work gets torn out.

Here's the problem in one building. This YMCA childcare centre's real public bid set has 54 sheets; two addenda replaced 20 of them. Both versions sit in everyone's inbox. On sheet P101, two of the three toilets in Toilet 105A quietly changed type. Order from the old sheet and two get ripped out.

Drawing Claim Checker lets a project team ask the current drawing set anything (typed, spoken, or filmed on a phone) and get the answer boxed on the actual sheet in about ten seconds. Search finds the current sheet using a revision index built from the title blocks. Two Gemini 4 models judge the crop in parallel and return the exact words they relied on, with bounding boxes. Then deterministic code checks their work: every quoted word must sit inside its box on the PDF's text layer, the sheet must be the current revision, dimensions are recomputed, and room outlines traced from the drawing's vector walls confirm the evidence is in the right room. If anything can't be proven, the answer is "unclear", never a guess. Gemini proposes; code disposes.

It also stops RFIs before they're sent when an addendum's Q&A already answers them (quoting the exact answer), writes a verified, plain-English addendum impact digest by trade in one long-context Gemini 4 Neon call, and shows the evidence in a 3D view built from the drawing's real geometry.

Measured on frozen held-out claims it was never tuned on (2026-11-21): 0 false alarms across 60 claims, and right every time it gave a verdict; held-out v2: 27/43 correct (95% CI 48–76%) (28% better than trunk tools). About 1,150 automated tests, including adversarial suites. Built entirely on Gemini 4, on real public documents.

## More from Gemini 4 NYC Hackathon

- [Sync](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/77)
- [AntiSlop](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/78)
- [Adaptive NPCs](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/79)
- [Chifles](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/81)
- [CrowdSite](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/82)
- [Worldbuilders](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/83)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/80. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt
