AI models such as GPT and Claude are powerful, but simply connecting an AI model to a CAT tool does not automatically produce professional, human-like translation.

We wanted to use AI differently: not simply to generate translations, but to build a complete translation workflow around the AI.

I've been using Trados and memoQ for years. After more than five years of developing, testing, and refining my own AI translation and QA tool, I learned that the AI model itself is only one part of the solution.

Here are some of the main challenges we needed to solve.

1. A robust SDLXLIFF/XLIFF parsing module

We wanted to process SDLXLIFF/XLIFF files directly.

A professional CAT file contains much more than text. It can include:

• Inline tags
• Segment status
• Comments
• Metadata
• Formatting information

During translation, these elements must be handled correctly.

The translation engine needs to preserve tags, maintain the file structure, update segment status when required, and generate a valid SDLXLIFF/XLIFF file after translation.

Although SDKs are available, we decided to develop our own parsing module so we could have more control over the workflow and handle the file structure according to our requirements.

2. AI needs human translation rules

Simply sending text to GPT or Claude and asking for a translation is not enough.

Human translators follow specific rules based on the client, project, terminology, style, and language pair.

We wanted to preserve those rules when using AI.

Our tool therefore uses specialized prompts and workflows for:

• Translation
• Editing
• Terminology
• IIT
• KTM
• TGM
• Other localization tasks

The goal is to give the AI the information and rules it needs to produce a more human-like translation, rather than relying only on the model's default behavior.

3. Translation Memory and glossary integration

Professional translation workflows depend heavily on Translation Memories (TM) and glossaries.

We wanted the AI to use these resources during translation.

The tool can use a user's existing TM and glossary to:

• Maintain terminology consistency
• Reuse previously approved translations
• Follow client-specific terminology
• Provide additional context to the AI

This is an important difference between simply using an AI model and building an AI-assisted translation workflow.

4. Our own fuzzy matching

Another challenge was matching new segments with existing TM entries and glossary terms.

I spent around three months developing our own fuzzy-matching module specifically for TM and glossary matching.

I did not rely on an existing Python fuzzy-matching library. I developed the matching logic myself so that it could be adapted to the requirements of localization workflows.

This gives us more control over how TM and glossary entries are matched and used during translation.

5. QA is part of the translation workflow

Translation does not end when the AI generates the target text.

The result needs to be checked.

Our tool supports QA checks including:

• IIT — Inconsistency in Target
• KTM — Key Term Mismatch
• TGM — Tag Mismatch

It can also work with Xbench QA reports and help fix issues identified in those reports.

The goal is to automate repetitive QA work and reduce the amount of manual correction required after translation.