Localization

Your app’s English text file in, the same file out in every language you pick, with the same keys in the same order. Two to four AI models each translate every line on their own, a judge picks or merges the best, and code checks every placeholder, plural and fixed word. Every line gets a confidence score, so you check the few lines that need it and ship the rest.

Asks for
  • Your app’s English text as a JSON file, up to 2,000 lines
  • The languages to translate into, up to ten at a time
  • Optionally, your industry, a sentence about the product, and words to keep or translate a fixed way
Gives back
  • One JSON file per language, with the same keys and order as yours
  • A confidence score on every line, and the lines to check picked out
  • Every model’s version of each line, with what it means in English
  • Any line changed in place before you download
Plan
Pro and Max

What it gives you

cancel.confirm   “Cancel your visit on <b>{date}</b>?”
  Claude Sonnet .. Annuler votre rendez-vous du <b>{date}</b> ?   ← judge’s pick
  GPT ............ Voulez-vous annuler votre visite le <b>{date}</b> ?
  Gemini Pro ..... Annuler la visite du <b>{date}</b> ?
  2 of 3 agree · judge 71 · placeholders OK · tone differs   → 74, check
One line, in French. Shortened.

How it works

One model asked to translate a JSON file gives one answer per line and no way to tell which lines are wrong. Here two or three models translate each line independently. Where they agree, the line is more likely right; where they do not, you see it.

A judge model that does not translate reads every version side by side, labelled A, B, C, never by name, and picks or merges the best. It scores each line and says in plain English why it chose that wording.

Then code checks what a model can get wrong quietly: every placeholder like {name} and %s, every tag, plural rules, the language’s own script, the words you fixed, and lines that came out far too long or too short. A failed check lowers the score, so it shows up as a line to check.

Who it is for

A PM shipping an app or a site in more than one language who needs the translations good enough to release, and wants to know which lines to have a native speaker look at.

Where it stops

JSON files only for now; YAML, .strings, XLIFF and .po are not read yet. Up to 2,000 lines and ten languages at a time, one translation running at once. A high score means the models and the judge agree and every check passed, not that a native speaker has read it.

Questions people ask

Does it keep my placeholders and plurals?

Yes, and it checks them in code rather than trusting the models. Every {name}, %s, tag and plural rule in the English must come back exactly. A line that loses one is sent back to be fixed once; if it is still wrong, it is marked red for you.

Why several models instead of one?

Because agreement is evidence. When models that translate on their own write the same thing, the line is very likely right. When they differ, the judge picks, and the line scores lower so you look at it.

How much does it cost?

The cost shows before you start, from the size of the file, the languages and the models you pick. Each language is charged as it finishes, for what its translation really used.

Run it on your own work

Localization needs an account, because it writes against what you give it and remembers what you have told it before. It is part of Pro and Max. Or see what the plans cost first.