Release-ready in every language.
Drop the files. Sign the estimate. Export only what a linguist and a live surface both accept.
No spend until you sign the estimate. Exact memory matches are free.
The localization operation,
delivered.
Same workspace. Now in every language you sell in.
Most teams live in two broken worlds: testers clicking through English, and translators working in files nobody opens in the real app. Flip KaDeep AI Studios to Globalization and one agent does both jobs, in one pass.
Checkout frameURL/checkoutYour glossary: 214 terms. Last year's translation memory: 38,200 segments.
en-US · iPhone · baseline
de-DE · iPhone · button cut off
Found by KaDeep AI Studios| Key | English | Deutsch | Source | Status |
|---|
A translation is a draft. A locale is a surface.
KaDeep AI opens the localized build in a real browser and, when you have a build, on a virtual device. It looks for truncation, overlap, off-canvas type, untranslated text and alt, RTL direction, lang, and layout drift. A vision pass reads the screenshot. What it finds is an ordinary defect, with a locale, in the same queue as QA.
- 01
Same journey, two languages.
The source and the target walk /checkout. The frames stay side by side. A clipped German CTA is SHOPPING-1440, not a comment in a spreadsheet.
- 02
Rules first, then a second model.
Placeholders, tags, plurals, numbers, dates, currency, empty targets, wrong script, DNT, and forbidden terms are deterministic. The MQM review always runs on a model that is not the translator.
- 03
CI speaks localization.
The suite that already gates the release can gate the locale.
loc translateloc qaloc validate --fail-on critical
Source English. Target Arabic, RTL. The nav mirrored. The price in SAR. Two untranslated alts filed.
You do not re-buy what you already accepted.
Exact and in-context memory matches apply without a model call. Fuzzy matches are hints. Glossary, DNT, and forbidden terms are constraints. Legal, finance, and medical stay in a human queue. Export is a checksum, not an email attachment.
Estimate
Words, images, and review, priced before the job. Memory is a credit on the bill, not a footnote. A campaign can start with a pilot locale.
Memory and glossary
Project or org scope. TBX and TMX in. The agent can propose a glossary from the content; nothing is active until you say so. We write decisions into memory. We do not write your content into a shared model.
Gates
A delivery is one asset in one locale: translating → in-review → needs-rework → approved → delivered. Critical flags block export. Tagged content needs a person. A minimum MQM score is a project setting.
Routing
UI, long copy, and low-resource pairs are different tasks. Hindi and Marathi default to a stronger tier. You can pin a model. Zero-data-retention providers can be required for the org.
You do not re-buy what you already accepted.
Exact and in-context memory matches apply without a model call. Fuzzy matches are hints. Glossary, DNT, and forbidden terms are constraints. Legal, finance, and medical stay in a human queue. Export is a checksum, not an email attachment.
Estimate
Words, images, and review, priced before the job. Memory is a credit on the bill, not a footnote. A campaign can start with a pilot locale.
Memory and glossary
Project or org scope. TBX and TMX in. The agent can propose a glossary from the content; nothing is active until you say so. We write decisions into memory. We do not write your content into a shared model.
Gates
A delivery is one asset in one locale: translating → in-review → needs-rework → approved → delivered. Critical flags block export. Tagged content needs a person. A minimum MQM score is a project setting.
Routing
UI, long copy, and low-resource pairs are different tasks. Hindi and Marathi default to a stronger tier. You can pin a model. Zero-data-retention providers can be required for the org.
Five markets. One checkout. One course. One app.
The first programs we built this mode for are the ones sitting in the studio today: a regulated checkout, a SCORM course, and a mobile build that has to survive German length and Arabic direction.
Checkout and regulated UI
Place order is not a slogan in Germany. It is statutory wording. The model drafts. The glossary holds the term. A reviewer signs. If the button clips, the defect is layout.
E-learning / SCORM
A zip becomes documents and segments. Placeholders and linked quiz keys stay intact. The player is in the studio. The export is the same package, per locale, with a checksum.
Product and mobile
JSON, ARB, iOS strings, Android XML, XLIFF, PO, RESX. CLDR plurals. Expansion budgets. RTL. The live device tab is the same tab QA already uses.
Who signs in
Localization manager
Locales, estimate, campaign, deliveries.
Linguist
Queue, terms, notes. Not the spend.
Reviewer
Legal and gated tags. The signature on the delivery.
QA lead
The same defects, the same dashboard, a loc facet.
Engineer
CLI, MCP, a failing gate in CI.
Formats we ingest on this page
Images: OCR to segments, then a render per locale, then a read-back. If the painted string is missing, it is a flag.
Start with one locale. Sign the rest when the surface is clean.
Create a localization project. Drop one asset. Pick Hindi or German as the pilot. The estimate is the first thing you see.
KaDeep AI Studios by Kadeep.ai. Testing and localization. Same studio.