Professional translation of technical books and documentation
You cannot simply run an eight-hundred-page book through a translator. A mistake here is not a matter of style: it is a token lifetime, a rate limit, a field name in a request. So we parse the document and settle the terms first, translate second, check third — and put it all back into the same file, with every bit of its formatting.
Book→ Terms→ Translation→ Checks→ File
A thesis and a ten-volume set of manuals go through exactly the same steps.
Not “text in another language”, but a book ready to work with: in its own format, with your own terminology and a clear list of the places worth re-reading.
DOCX, EPUB and HTML are rewritten in place: formatting, images, tables, the table of contents and styles all stay where they were. Nothing to lay out again.
Terms are settled before translation and hold in every chapter and every volume of a series. Not “housing” in chapter 2 and “casing” in chapter 7.
Every segment passes six checks. Anything doubtful is flagged and collected into its own queue — there is no need to proofread the book front to back.
How it works
What separates this from ordinary machine translation is what happens before and after the translation itself. That is where the mistakes live that a reader finds first.
PDF, DOCX, EPUB, HTML, Markdown or plain text. The document is broken into segments — headings, paragraphs, list items, table cells — and its structure is remembered.
The system collects what repeats through the book and shows it as a list. You decide once; while undecided words remain, translation does not start.
An unclear word comes back with a reference: a definition, the translation accepted in the field, the expansion of an abbreviation and the addresses of the sources. What is found stays in your catalogue.
The model receives the segment, the paragraphs around it, the section heading and the glossary terms — as a requirement, not as a hint. Repetitions and memory matches are not translated again.
Numbers, units, placeholders, glossary usage, abbreviation expansion. An edit to one segment propagates to all of its repetitions; a human signs the work off.
A finished file in the same format. An unfinished document is never handed over silently — you see how many blocks were left as source text.
Terminology
A word left to the model’s discretion will be called several different things across forty places in the book — and the text will still read smoothly, so only the reader will notice. That is why the list of everything that repeats is collected up front and settled once.
The risk that the price of the underlying asset and the price of the derivative diverge while hedging: the basis moves, and the hedge stops being complete.
The source is saved together with the decision
Who needs this
The task is the same: terms must hold, numbers must not drift, and the file must open exactly as it was. Only the volume differs, and how many people work on it.
A textbook, a monograph, a paper, a thesis, a single chapter.
Volumes of documentation, series of manuals, product localisation.
Quality
The checks do not judge style — they catch what is expensive in a technical book and invisible while reading it.
Repetitions inside the book and matches with earlier orders are filled in without calling the model. The same thing reads the same way.
Fix a warning once and you have fixed it in all forty places — except the ones a human has already accepted.
The address where the term was looked up is saved together with the decision. An argument about a term does not start from scratch a month later.
Formats
EPUB is not text, it is an archive: cover, styles, fonts, images, manifest and table of contents. You cannot rebuild that out of translated paragraphs, so the original is rewritten in place and everything else is carried over untouched.
Any format can be collected as plain text, as Markdown or as a new Word document — to proofread in your usual editor, for instance.
Cost
Every run reports back: how many segments were covered by memory and repetitions, how many went to the model and what that came to. Tokens are stored, not money: prices change, and what was spent on this book is a historical fact.
The same warning occurring forty times is translated once — not to save money, but so that it reads the same way throughout the book. Paying for it once is a pleasant consequence.
No, and it does not try to. The machine covers what a human does slowly and with mistakes: keeping terms identical across eight hundred pages, reconciling numbers and units, filling in repetitions. The decision on anything doubtful stays with the human, and the work counts as done only once they have accepted it.
They are carried over unchanged — that is a separate translation rule and a separate check on the output. Notation (σ, σᵤ, LT1) can be entered into the glossary as entries of its own: a substituted subscript does not read like a typo, it changes what the formula means.
Yes: CSV, TSV, TBX and working terminology registers in DOCX. What you upload never overwrites what a human entered — those rows end up in the skipped list with a reason instead of disappearing silently.
The longest part is not the translation, it is deciding on the terms — that is human work, and speeding it up at the cost of quality is pointless. The translation itself runs in batches of segments; repetitions and memory matches are not recomputed at all.
Organisations are isolated from one another at the data model level, not in interface settings. The external reference lookup is off by default — it goes out to the internet. The translation provider is a setting, including a model you host yourself where the text must not leave your perimeter.
Yes. A document is translated in parts, and a repeat run does not touch what you have already edited and accepted. Translation memory accumulates as you go — the next chapter and the next book come out cheaper and more consistent.
Send us a book or a manual — we will set up a workspace, parse the document and show you the list of terms before a single cent is spent.