Flagship service

Translation that holds its terms.

Technical catalogues, documentation and compliance text translated end to end by machine — with terminology governed by a glossary rather than guessed by a model, and quality checked automatically on every block. No human review stage.

Generic machine translation guesses.

For everyday prose that is fine. In technical and regulated fields it is expensive, because the failure is invisible until it reaches a customer or an auditor.

The same component is called three different things across three documents. A loanword that the entire industry uses in English gets "helpfully" translated into something nobody searches for. A tolerance, a thread standard or a material grade is rendered as an approximation. Each error is small; together they make the output untrustworthy.

So a human is put in front of it. That person re-reads everything, fixes the same handful of terms again, and the organisation concludes that translation simply costs what it costs.

The cost is not the translating. It is the checking — and the checking exists because nothing in the process guarantees the same term resolves the same way twice. That guarantee is what we build.

Four stages, one governed glossary.

Terminology is resolved before any full text is translated, so the glossary constrains the translation rather than being reconciled against it afterwards.

Term selection Finds domain terms Term translation Builds the glossary Full text Glossary enforced Automated QA Checks every block Findings consolidate into the glossary overnight Segmentation is structural, by block — not sentence by sentence
Fig. 1 — Translation pipeline. Glossary commits happen only via nightly consolidation, never mid-run.
01

Term selection

The source corpus is scanned for the vocabulary that actually carries risk: component names, industry loanwords, standards references, material and thread designations. These are the terms where a wrong choice propagates across thousands of documents, and they are almost never in a general dictionary.

02

Term translation

Each selected term is resolved once, per language, with its rationale recorded — including the decision to leave a term untranslated where the industry uses the English form. The result is a governed glossary rather than a lookup table: every entry has a reason attached, so it can be reviewed, challenged and kept stable as the catalogue grows.

03

Full-text translation

Only now is the body text translated, with the glossary enforced as a constraint. Segmentation is structural — whole blocks rather than isolated sentences — so the model keeps the surrounding context that determines how a term should behave, instead of translating each line blind.

04

Automated review

Output is checked programmatically against the glossary and against the source structure: term adherence, placeholders, numbers, units, formatting. Findings are written to an evidence store and consolidated into the glossary on a nightly cycle — so the system improves between runs without anyone editing text by hand.

Not a translation API. Not a translation platform.

Raw machine translation is a component; localisation platforms are workflow tooling that still assumes people in the loop. The gap between them is where this sits.

Raw MT API Localisation platform Icarops
Terminology Per-request glossary at best; no governance Term base exists, maintained manually Governed glossary, updated by the system
Quality assurance None QA checks, resolved by a human reviewer Automated checks with an evidence trail
People required Someone to review the output Reviewers, and someone to run the tooling None in the running loop
Consistency across runs Not guaranteed Depends on reviewer discipline Guaranteed by construction
Cost driver Per word, plus the review you still pay for Seats and volume The pipeline, not the headcount

Comparing on price per word compares the wrong thing — the review stage is where the cost usually sits.

Case

An industrial fluid-control distributor

A European distributor of industrial fluid-control components runs a multi-language product catalogue where terminology is unforgiving: valve types, connection standards, thread designations and materials all have established industry forms that differ per language — and several that must deliberately stay in English.

Product content is translated from a single source catalogue into five languages through the pipeline above. Terminology is resolved once and enforced everywhere, so a component is described the same way in every language, on every page, on every run. Corrections found in review are consolidated into the glossary overnight rather than re-applied by hand.

175,000
Products in scope
3 million
Words per month
3.5 FTE
Of review work removed

Content we handle

Product catalogues and PIM content, technical documentation, datasheets and specifications, compliance and regulatory text, category and marketing copy where terminology still has to hold.

Getting started

A new domain begins with the glossary, not the text. We build the term layer from your existing content, agree the entries that matter, and only then put volume through the pipeline.

Have a catalogue that needs to hold its terms?

Tell us the domain and the languages. We will tell you whether it is a fit.

Get in touch