# Colophon

- **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:** [Rodrigo Conde](https://cerebralvalley.ai/u/rconde)
- **GitHub:** https://github.com/rtgconde/colophon
- **Demo video:** https://youtu.be/BdbA1Wr3TI0
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2975

Every picture in the Museum of Modern Art's free exhibition catalogues, found, read and indexed; every catalogue identified from its own pages and checked against the museum's records, which it corrected once. Colophon makes a scanned library behave like a catalogued one.

Try it: https://colophon.rconde.tech is a public copy of everything Muse Spark read (284 catalogues, 22,297 pages, 45,489 figures): search, the works reproduced across catalogues, the catalogue list and the page view with every figure boxed. It is a frozen snapshot with keyword search; the semantic search in the video runs on the local index.

MoMA gives 757 out-of-print catalogues away as PDFs, 77,750 pages from 1929 to today, but a scanned book has no index of its pictures and often no record of what it is. The same is true of the thousands of scans in my own library, which is why I built this.

Two jobs. First, find and read every picture on every page: SAM 3.1 boxes every figure (93,527 of them, 5 h 12 min on three GPUs), then Muse Spark reads each page with the boxes drawn and numbered and writes one record per figure: what it shows, the caption as printed, artist, title, year, medium, credit. That record is embedded and indexed, so you can search for what you remember ("a woman reading a letter", "water tower") and open the page with the figure outlined. A Works view joins the same work across catalogues: 796 works are reproduced in two or more.

Second, identify every book from its own pages: Muse Spark reads the title pages and the verso and returns title, author, publisher, year and ISBN with the page each answer came from; the ISBN is checked against Open Library and Google Books. Scored against MoMA's own records for all 757 books: title 94 %, year 90 %, publisher 86 %, ISBN 92 %, 7.5 s per book, $0.031 per book. It caught a wrong date in the reference records.

Everything in the repository was built during the hackathon: the page pipeline, the SAM workers, the reader, the index and search page, the identifier and its scorer. The library itself and the machines it ran on are my own.

## More from Meta Global AI Developer Hackathon

- [MATRIX-code](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2972)
- [SheRules](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2973)
- [slope coach](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2974)
- [Helpa](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2976)
- [DevTeam AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2977)
- [NearMiss](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2978)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2975. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt