AI translation and localization at scale · Arabic and nine more languages
AI translation that shows its work.
From your first batch to your largest program, every file runs the same seven-step AI cycle and comes back with a quality report: what was checked, what changed, and why.
Reset your password from the account page.
Setzen Sie Ihr Kennwort Passwort auf der Kontoseite zurück.
- Found
- AgentTerminology check: “Kennwort” is not your approved term for “password.”
- Challenge
- Devil’s advocateBoth words are correct German. Is this a real error?
- Ruling
- JudgeYes. Your glossary sets “Passwort.” Change it. Severity: minor.
- Human review
- None
- Quality gate:
- Passed
Your program grows; the method doesn’t thin out. One large batch or release after release, from your files or inside your TMS.
How we run large programs- New, matched, and repeated words counted at intake
- Unchanged text isn’t translated again
- A service level for each content type
Watch one file go through the cycle
One file, seven steps, in a video of about a minute: the checks, the devil’s advocate, the judge, and the quality report that comes back with the file.
Read the transcript
AI translation is often a black box. Ours isn’t.
Every file goes through the same seven-step AI cycle: intake, terminology, translation to your glossary and rules, then separate checks, each blind to the others.
Then a devil’s-advocate agent challenges every finding, and a judge agent makes the ruling. No change rests on one agent’s word.
Then the quality gate. And every file comes back with a quality report: what was checked, what changed, and why.
From your first batch to your largest program, no step is skipped.
Locstars. AI translation that shows its work. locstars.ai.
For teams that publish more than they can read
When your content goes out in languages your team can’t read, a translation is not enough. You need a record of what happened to it.
Content and localization leads
AI translates a growing share of your content, in languages your team can’t read. Every file comes back with a report that shows what the cycle found, what changed, and why.
Teams growing into new markets
Raw, unchecked AI output isn’t something you publish. Our cycle checks, challenges, and rules on every file before it ships, batch after batch.
Translated. Challenged. Judged. Then released
The same seven-step cycle runs on every file, at every level. Here it is in four stages. Nothing is released on one agent’s word.
-
Translated
An AI agent translates against your glossary, translation memory, and style rules.
-
Challenged
Separate agents check accuracy, terminology, grammar, and style, each blind to the others. A devil’s advocate then challenges each finding and asks for its evidence.
-
Judged
A judge agent weighs each finding against its challenge, rules on it, and records the reason.
-
Optional Human add-on
An add-on, when you ask for it. A linguist joins after the judge’s ruling and before the quality gate: + Human check, + Human edit, or + Full review.
-
Released
An automated quality gate checks the delivered file. If it fails, the file doesn’t ship.
Don’t take our word for it. Read the report
You don’t need to read the target language to know what happened to your file. The report lists what the cycle checked, every issue it found, how each one was settled, and the level that ran.
- Job
- Help article, English → French, fictional client
- Level
- AI
- Human review
- None
- Steps run
- All seven, including the devil’s advocate and the judge
- Issue
- Accuracy · Critical
- Source
- Do not share your codes with anyone, including our support team.
- AI translation
- Ne communiquez vos codes à personne, sauf à notre équipe d’assistance.
- Found by
- Accuracy check: “including” became “sauf” (except).
- Devil’s advocate
- The warning still applies to everyone except support. Major, not Critical?
- Judge
- Critical, not Major. Readers could give their codes to anyone claiming to be support and lose control of their account. Fix before delivery.
- Delivered
- Ne communiquez vos codes à personne, pas même à notre équipe d’assistance.
- Quality gate:
- Passed
The full cycle at every level. Human review is an add-on
The AI level is a complete service: the full seven-step cycle and its quality report on every file. Human review is an add-on, per content type, when you ask for it.
-
AI
The full cycle and its quality report on every file.
Human review: None
-
+ Human check
A native linguist checks the segments the cycle flagged, plus a random sample.
Human review: check
-
+ Human edit
A linguist edits every segment.
Human review: edit
-
+ Full review
A linguist edits every segment, then a second linguist reviews it.
Human review: full review
Every report states the level that ran: Human review: None, check, edit, or full review.
Your customers don’t all speak the same Arabic. Your translation shouldn’t either
Arabic dialects are core to what we do. We translate into Modern Standard Arabic (MSA) by default, and into Gulf, Egyptian, Levantine, Maghrebi, Iraqi, or Sudanese Arabic when your audience speaks it. Every report names the variety we wrote in.
Left unchecked, AI tends to slip back into MSA or mix two dialects in one sentence. So the cycle runs a dialect-consistency check on each file written in a dialect.
- Modern Standard ArabicThe shared written standard, and our default.
- GulfSaudi Arabia, the UAE, Kuwait, Qatar, Bahrain, and Oman.
- EgyptianEgypt, and widely understood across the region.
- LevantineLebanon, Syria, Jordan, and Palestine.
- MaghrebiMorocco, Algeria, and Tunisia.
- IraqiIraq.
- SudaneseSudan.
| MSA | أريد أن أرى العرض الجديد الآن. |
|---|---|
| Gulf (Saudi Arabia) | أبي أشوف العرض الجديد الحين. |
| Egyptian | عايز أشوف العرض الجديد دلوقتي. |
| Levantine (Lebanon, Syria) | بدّي شوف العرض الجديد هلّق. |
| Maghrebi (Morocco) | بغيت نشوف العرض الجديد دابا. |
| Iraqi | أريد أشوف العرض الجديد هسّه. |
| Sudanese | داير أشوف العرض الجديد هسّع. |
Same meaning, a different word for “want” and for “now” in almost every line. One common spelling is shown; spelling varies.
Ten languages, including Arabic, where we know AI’s weak spots
Each language pairs with English in both directions, and Arabic pairs with each of the others. Each one runs the same seven steps and comes with a quality report.
AI output quality varies by language pair and content type. Before you commit, we can run a sample on your own content.
- EnglishThe hub: every language pairs with it
- العربيةArabic · MSA and six dialect groups
- FrançaisFrench
- EspañolSpanish · Spain and Latin America
- DeutschGerman
- 简体中文Chinese (Simplified)
- TürkçeTurkish
- Português (Brasil)Portuguese (Brazil)
- ItalianoItalian
- 한국어Korean
One cycle behind every service
AI translation and localization is the core. Every other service runs on the same seven steps, with a report on every delivery.
AI translation and localization
Large batches and ongoing programs, from product UI to legal and medical text.
Continuous localization
Websites, apps, catalogs, and help centers, in step with your releases.
SEO and AI-search localization
Keywords researched in each market, not translated from English.
Scanned PDF rebuild and translation
Scans in. Editable files out, in the original layout.
Subtitles and captions
AI subtitles and captions, with timing checked like the text.
Building AI that speaks Arabic? See AI assistant language testing. Already translating with AI? Free check.
Start with one file. Judge us on the report
Request a quote for one file, or tell us how much content you have, and whether it’s a one-off batch or ongoing updates, and we’ll scope the whole program. Before any work starts, we agree on the level for each content type and what you get back.
Want to know where your content goes? Read how we handle your data.