How I cut book translation from 15 hours to 1

Michaelshof Sammatz runs a library that buys books no longer in print and scans them to preserve in its own collection, in languages that include Russian, Spanish, Swedish, Portuguese, English and French. Before, translating one of those scanned books meant roughly 15 hours of a freelancer’s work stretched across weeks of back and forth with the client. I built a system that takes the scan, translates it chapter by chapter with an LLM, and hands back a fully formatted Word document in about an hour of processing, for roughly €15 in tokens per book.

The situation

The library’s scanner is a Treventus, and every scan comes out as a PDF, whether it’s a book or a magazine. Before this system existed, translating one of those PDFs into a usable book meant sending it to a freelance translator, going back and forth with them on questions and revisions, and waiting weeks per book. Books needed to end up in Word format, laid out correctly, so the library could either print them or just check them over. Translation was one of four things running at Michaelshof Sammatz that did not actually need a person doing the work by hand.

The build

The system watches a shared Google Drive folder that the library uploads scans into, checking it every few minutes. When a new PDF lands, it reads everything inside the scan, translates every chapter with an LLM, and puts the book back together so the output matches the original design: the same structure, title page, and back matter the scan came in with, not just a dump of translated text. If a book is expensive enough to translate that it is worth flagging, the system posts to a Telegram channel first with the estimated cost, so translation waits for approval before it starts. Once a book is done, it lands in two places: the finished document goes into the Telegram group along with what it actually cost to translate, and a copy goes back into Google Drive.

The result

The system has been running since it was delivered, and nobody at the library has needed to touch it. A scan that used to cost a freelancer 15 hours of work spread over several weeks now comes back as a finished, correctly formatted book in about an hour of processing, for around €15 in LLM tokens.

The library still uploads scans the same way it always did. The only difference is what comes back.

FAQ

How do you automate translating a scanned book into another language?

A scan goes into a shared folder, gets read chapter by chapter, and is translated with an LLM before being reassembled into a formatted document. The output keeps the same structure as the original scan, so it comes back ready to use rather than as raw translated text.

Can AI translation preserve a book’s original formatting?

Yes, if the system is built to read the book’s structure and not just its text. This system rebuilds each book with the same layout, title page, and back matter as the original scan, which is what makes it usable without manual cleanup afterward.

How much does it cost to translate a book with AI?

For Michaelshof Sammatz’s library, it costs roughly €15 in tokens per book, compared with a freelance translator’s fee and weeks of turnaround for the same book before this system existed.