Case file · 01
Clear Cut Translate
A private AI translation system for the technical documentation of a German hydraulics manufacturer — built around the customer’s own language, not a generic engine.
The living document
The product’s public landing is one continuous scroll act: a document enacts the entire pipeline — extract, match, translate, verify, reassemble. This is the real landing, embedded live.
The full scroll journey is best on a desktop screen.
Live demo of the real product — identifying customer details are anonymized.
The problem
For an industrial manufacturer, translating technical documentation is slow, expensive and risky: generic tools break DOCX layouts, miss house terminology, and every uncaught error ships to customers inside datasheets and manuals.
Meanwhile two thousand already-translated documents sat in the archive — the company’s entire technical language, encoded in pairs nobody could query.
The approach
Instead of sending everything to a model, the system first decides what is already known.
- 01
Translation memory from the customer’s own corpus
Every segment is matched against the archive first — exact and semantic matches carry over untouched, so the company’s established wording always wins.
- 02
AI only for the remainder — glossary enforced
What the memory can’t answer goes to the model with the house terminology attached, then through staged verification: terminology, numbers, completeness.
- 03
Format-true DOCX round-trip
A patch-based pipeline edits only the text inside the original file, so layouts, fields and numbering survive byte-for-byte — export, translate, reimport.
Where it stands
In build for its first customer. The skeleton runs end to end — upload to reassembled DOCX — and the landing above is the product’s own front door, built with the same care as the pipeline behind it.
- 2,000+
- documents in the translation-memory corpus
- 5
- pipeline stages, each independently verified
- 1:1
- DOCX layout round-trip as the acceptance gate
References available on request.
Need something built like this?
Clear Cut Translate started as a conversation about a slow, error-prone process. If you have one of those, we should talk.
Start a conversation