Restoring a scanned firm history book, and what my spot check caught
The firm's decades-old history existed only as a 100-page scan with a poor text layer, and the automated OCR cleanup looked clean on the first pass. It wasn't. A page-by-page check against the original found that the hardest parts, the tabular sections, had been silently scrambled, which is the entire reason that check existed.
Category
Document restoration and verification discipline.
What I was trying to accomplish
Turn a decades-old, 100-page scanned firm history book with a poor text layer back into something usable: searchable, transcribed, and available as an accurate, editable edition that kept the feel of the original.
The data challenges
Automated OCR cleanup looked clean on a first pass, but had actually scrambled or dropped the hardest parts of the book, something only a page-by-page spot check against the original scan caught.
The result
A corrected transcription, a searchable PDF, and a 68-page editable Word edition that follows the original layout, with the damaged tabular sections rebuilt from scratch and verified page by page, and only minor word-level OCR typos remaining.
Technical details
Tools: Claude Cowork, Tesseract OCR, Python (text cleanup), Node.js docx, LibreOffice rendering for visual checks
The firm had a decades-old printed history of itself, available only as a 100-page scanned PDF with a poor text layer. The goal was to make it usable again: a searchable PDF, a clean text transcription, and an accurate, editable Word edition that kept the feel of the original.
Claude re-ran OCR on every page with Tesseract, cleaned the text, and produced a Markdown transcription and a searchable PDF before moving to Word. The first pass looked clean. But a spot-check against the scan found that a table was missing, and the complex parts of the book, a 61-row roster table, two fee schedules, and an indented lineage timeline, had been scrambled or dropped by the automated cleanup.
Claude made verification the core of the work from that point: it cut the damaged blocks out, rebuilt them as real Word tables and formatted lists in a scripted two-pass build, compared the result to the scan page by page to confirm no narrative text had been lost, and rendered pages to images to check the layout visually. The page-by-page audit found the narrative intact and the damage confined to the three tabular sections, all rebuilt.
Pull up a chair.
Let's talk through where things stand, and where they could go.