Notes on the state of archive digitisation and OCR, and on what is changing inside Tetrak.
Two threads run through this section, and they are the same thread seen from either end.
The first is the field: what OCR can and cannot currently do with the material that actually sits in archives — the linen postcards, the multi-column trade journals, the hand-lettered posters — and how the people who look after those collections are getting text out of them today.
The second is this project. Where the routing has been changed and why, what the quality gate is now catching that it missed before, and how the Armenian EasyOCR model is coming along. Progress reports, including the ones where the number went the wrong way.
Nothing here restates the benchmark. The research is where the measured comparison lives, and it is regenerated from the harness rather than written; these are the arguments and the working notes around it.
Ten human-proofread encyclopedia pages, ten OCR configurations measured against them, and the first honest numbers from a recogniser we are training ourselves — with the code, data and weights all published in the open
Opening this section with the two things it is for: the state of OCR for real archival material, and the working notes from building a tool for it.
New articles turn up in the RSS feed.