I often see people say that AI is not good enough for professional translation or QA.
I understand why they think that—if you're only using GPT or Claude directly, they're right. On their own, they cannot replace a professional translation workflow.
I've been using Trados and memoQ for years, but I don't rely on them for the actual translation anymore. Instead, I developed my own AI-powered translation and QA tool after more than five years of testing and refinement.

I developed an AI-powered translation tool that can directly process SDLXLIFF/XLIFF files and provide fast, high-quality translations while maintaining the original file structure.
The tool supports Trados and memoQ files, allowing users to select the required language pair and apply a glossary to maintain consistent terminology throughout the translation process.
After translation, it also provides automated QA functions, including Inconsistency in Target (IIT), Key Term Mismatch (KTM), and Tag Mismatch (TGM) checks, and can apply Xbench AutoFix to automatically resolve common QA issues.
One of the key features is that it preserves tags during translation, which helps avoid formatting and structural problems that often occur with automated translation workflows.
These are the main features currently available, and I am continuously improving the system with more capabilities.

Here are some of the key challenges I had to solve:

1. A robust SDLXLIFF/XLIFF parsing engine

Professional CAT files contain tags, statuses, comments, and other metadata that must be preserved.
The parser must correctly read the file, keep every tag intact, update segment statuses, and generate a valid SDLXLIFF/XLIFF file after translation.
Although an official SDLXLIFF SDK exists, I found it more flexible and easier to build my own parsing module tailored to my workflow.

2. Human translation rules

Every translator and every client has their own translation guidelines.
To achieve human-like translations, these rules must be incorporated into the prompts instead of relying solely on the default behavior of GPT or Claude.
My tool includes specialized prompts for translation, editing, IIT, KTM, TGM, and other localization tasks.

3. Translation Memory and Glossary integration

A professional translation system must use Translation Memories (TM) and glossaries during translation.
This ensures terminology consistency and allows previous approved translations to be reused instead of generating everything from scratch.

4. Custom fuzzy matching

I spent three months developing my own fuzzy matching engine for Translation Memories and glossaries.
I didn't rely on any existing Python libraries—I designed and implemented the matching algorithm from scratch to better fit localization workflows.

5. Professional QA automation

Translation quality doesn't end when the text is generated.
My tool performs important QA checks such as IIT, KTM, and TGM, similar to Xbench. It can also automatically fix issues detected in Xbench QA reports, significantly reducing manual post-editing.

I've been developing and testing this system for more than five years, continuously fixing edge cases, improving translation quality, and refining the workflow. I'm still improving it as I discover new challenges in real-world localization projects.

Thanks