Local AI for technical translation agencies

“Housing” in two projects

Say that in client A's approved glossary the word “housing” has the Polish equivalent “obudowa”. In client B's project, in the context described there, “korpus” was approved. The model prepares a draft for B and uses “obudowa”. The sentence sounds natural, and the translator still has to correct it to match this project's resources.

In the proposed view the translator sees which project a segment belongs to and can open the right terminology entry next to the source and the draft. They check the context, change the term and approve the text for the next step. The model helps prepare a proposal; judging technical correctness stays with the people who review the translation. When the glossary does not settle a given use, the question goes back to the person responsible for terminology.

What the translation memory and the termbase already give you

A translation memory stores earlier translations for reuse. A termbase holds the agreed equivalents of words and phrases. The memoQ documentation describes using both resources during translation.

If corrections happen because a project uses the wrong termbase, or approved changes do not reach the team's resources, it is worth tidying that work in the CAT tool first. Adding a model would leave the problem in every following job. Existing memory matches and term suggestions may be enough for repetitive segments.

Local AI is worth considering for segments that need new wording or editing, when the translator wants a proposal that takes the client's resources into account. Switching projects can mean choosing the right material for the app. It does not automatically require separate fine-tuning of the model for each client.

Where the text goes while you work

Say the client has allowed the text to be processed only in a defined environment at the agency. A local model can run on a computer or a server in that environment. Where it runs does not, on its own, describe the document's path through the whole app.

During integration you need to check where files and saved drafts stay and whether the app's activity logs contain pieces of text. A CAT plug-in or an extra service may send content outside that environment.

The person responsible for IT compares the actual flow with the requirements agreed for the project, sets access to the material and decides how support requests that need a look at the text are handled. These arrangements also cover later updates to the integration.

Do the proposals make the translator's work easier?

The comparison should cover the language pair and the documents the team really works with. The translator judges how many changes a proposal needs, whether it keeps the meaning of the source and whether it uses the right client's terminology. Also count the waiting time on the hardware you use and moving text between windows. A shorter correction of a single sentence may not make up for an awkward process across the whole job.

The agency can compare this work with its current way of translating in the CAT tool. The decision needs reviewed proposals and the translators' comments, including examples where they rejected the draft. This shows for which material the integration is useful and what still needs improving before wider use.

The proposed scope for a translation agency

As an AI app, Syntalith can prepare a local workspace linked to the client's resources and an agreed way of passing checked text to the CAT tool. The first scope can cover a chosen language pair, a connection to the project resources available and a view for comparing the source with the draft. How the integration works depends on what your software allows and on the approved data flow.

The agency's team names the corrections that keep coming back and the person who will assess the translations and approve the terminology resources. Together we set the material for the trial, the working environment and later maintenance of the connection to the CAT tool. That way you can look at the proposals in the translator's daily work before widening the scope.

Describe a terminology correction that comes back job after job, the CAT tool you use and where the text may be processed. We can start the conversation from that without sending confidential documents. We price a private model after a free call about the task, the data and the hardware. An app that uses the model starts from €3,000 net (pricing).

What a local model for technical translation costs and when it pays back

Where to start

In the free process review we choose one task and 20 to 100 documents the team has already worked on, and within 2 working days you get a written summary with a price. We roll out the local model as part of private LLMs, and the app around it as an AI app.