Localization QA Agent

Problem

Translations come back from a vendor or a model, nobody on the team reads the target languages, and errors ship to exactly the markets where nobody will report them - they will just convert worse.

Solution

An agent that reviews each string against the glossary, the locale's register, the placeholder contract and the space it has to fit in, and raises only the issues a human reviewer would have changed.

Tech stack

Required integrations

• TMS (Phrase, Lokalise, Crowdin)
• Source repository for string context
• CI, so a failing check blocks a release

Key features

• Placeholder and ICU plural validation, which is deterministic and catches real bugs
• Glossary enforcement per locale
• Register and formality check against the market
• Length overflow prediction against the component the string renders in
• Only reports what a reviewer would change, with a suggested fix