# Storey

- **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:** [Tarkeshwar Narayan Sharma](https://cerebralvalley.ai/u/tns)
- **GitHub:** https://github.com/tns15june/storey
- **Demo video:** https://youtu.be/kafwxdt-l20
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/863

Storey is an AI plan review assistant for city building departments. Before a building permit is issued, a plan examiner reads the whole drawing set by hand to check that people can get out of the building: that every stair continues floor to floor to an exit, that exits are wide enough for the occupant load, and that the set agrees with itself. In 2021 Portland's city auditor found that more than 80% of standard permit plans needed corrections, and each correction cycle costs the applicant weeks. Seattle's own AI pilot passed 6.4% of the checks a human examiner later sent back, and nobody looks again at what was passed.
An examiner uploads a PDF set. In about three minutes, for about half a dollar, Storey lists every sheet, follows every stair across every floor in a stair-by-floor grid, and raises a correction list. Each item gives its sheet, its place on the drawing, the value read, the rule for Houston, Philadelphia or Seattle, and a crop. Storey also lists what it could not check and drafts the comment letter. On a resubmittal it checks each earlier comment at its place and lists changes that no revision cloud marks. It has no pass state: whatever it cannot read shows as Cannot tell, and the examiner decides every item.
On 33 evaluation sets run once on code frozen by hash, Storey raised 184 of 184 planted items on generated sets, against 81 of 177 for Muse Spark asked directly, and showed 1 wrong pass in 16 held-out stair cells. On real sets audited by hand it found 6 of 121 items, against 15 for Spark asked directly, because most of those items are kinds its 17 checks do not cover yet. That is the next work.

## More from Meta Global AI Developer Hackathon

- [Pixel & Sprout](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/860)
- [Bravon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/861)
- [SecureCoating Meta](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/862)
- [Valorm Technologies](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/864)
- [Areazero.ai](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/865)
- [WireProof](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/866)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/863. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt