Skip to content
Beta

Translating in Our Pipeline With STACKIT AI Models

Last updated on

Stackit LogoStackit Logo
STACKIT

Translating in Our Pipeline With STACKIT AI Models

What we tested before going live and how two scheduled jobs translate merged pull requests with STACKIT AI models, checked in code and approved by a person.

PLAN

A Draft in Seconds Is Not Yet a Translation

A model writes fast. It can also break a link or invent a sentence.

The gap Three figures. About 140 English assets have no German version. Eleven test pages were translated in 6 minutes. One person checks every page. about 140 English assets without German 6 min for eleven test pages 1 person checks every page
About 140 English assets without German, 6 minutes for eleven test pages, one person who checks.
STACKIT LogoSTACKIT Logo
AI-Assisted Translation: Flow and Human Review STACKIT · Blueprint Open asset ↗

Every page in this framework exists in English. A German twin exists for some of them. Writing a twin by hand is slow, so the gap grows with every pull request. Today about 140 English assets have no German version.

A model can write a first German draft in seconds. That alone is not a solution. A draft can break a link or invent a sentence. Nobody notices until a reader does. The pipeline below keeps the speed and shrinks that risk. What the code cannot judge, a person checks.

AUTO

One Tick Starts the Whole Route

The author decides. STACKIT AI translates once the change is live. The author reviews the result.

One tick to a reviewed translation Five stages. 1: the author ticks the box. 2: the change goes live. 3: STACKIT AI translates. 4: a new pull request runs the same checks. 5: the author reviews and Core approves. 1 Author ticks the box 2 The change goes live 3 STACKIT AI translates 4 New pull request, same checks 5 Author reviews, Core approves
Five stages from one tick to an approved translation.
PLAN

Five Tests Before the First Real Run

We tested on our own pages before a single real pull request was translated.

Five tests before the first real run Five tests on our own pages before the first real run. 1: whole pages or pieces, sent whole 2 to 4 of 11 pages broke, in pieces Qwen3.8 kept all 11. 2: the cutter alone rebuilt 222 pages byte for byte. 3: eight models on the same 11 pages, Qwen3.8 came closest to our German. 4: the guard on 8,056 pieces of our own translations raised 11 false alarms. 5: a dry run over 150 closed pull requests opened none without a tick. Then the first real run: pull request 377 was translated as 378 at the next hourly run. 1 Whole page or pieces whole, 2 to 4 of 11 broke; in pieces, Qwen3.8 kept all 2 Cutter without a model 222 pages taken apart and rebuilt byte for byte 3 Eight models same 11 pages, Qwen3.8 came closest to our German 4 Guard on human text 8,056 pieces of our translations, 11 false alarms 5 Dry run on history 150 closed pull requests, none opened without a tick → First real run #377 translated as #378 at the next hourly run
Five tests on our own pages: whole pages or pieces, the cutter alone, eight models, the guard and a dry run over 150 pull requests. Then the first real run.
STEP

Small Pieces Keep Pages Intact

Sent whole, 2 to 4 of 11 pages broke. In small pieces our chosen model kept all 11 intact.

A command on its way through the translation Four steps in two rows. The English sentence holds a command. The AI sees a marker instead of the command, returns the German sentence with the marker, and our code puts the command back. 1 English Run pnpm build first. 2 AI sees Run ⟦1⟧ first. 3 AI returns Führe zuerst ⟦1⟧ aus. 4 Finished Führe zuerst pnpm build aus.
A command becomes a marker and comes back after the translation.
STACKIT LogoSTACKIT Logo
Pieces and Guard Rails: Keeping a Model in Its Lane STACKIT · Policy Open asset ↗
  1. Split: a parser reads the page and cuts it into pieces of plain text.
  2. Protect: links, inline code, components and image targets become placeholders such as ⟦3⟧. Code blocks and imports never leave the file.
  3. Translate: the model sees only numbered pieces with placeholders.
  4. Correct: a second model improves the pieces that need it.
  5. Restore: code puts every placeholder back at its original place.

The model does not see a link, so it has no chance to change it. The rules for German links, such as the de/ prefix and German anchors, are applied by code as well.

BASE

Open Models Run on STACKIT AI Model Serving

Switching a model means changing one name. The text stays on STACKIT.

Open models on STACKIT AI Model Serving A band for STACKIT AI Model Serving holds two models in use: Qwen3.8 translates and Gemma 4 corrects. A dashed card says more open models can be used by changing one name. STACKIT AI Model Serving Qwen3.8 translates Gemma 4 corrects More open models switch by changing one name
Qwen3.8 translates and Gemma 4 corrects. More open models run on the same endpoint.
STACKIT LogoSTACKIT Logo
Translating With STACKIT AI Models: What We Measured STACKIT · Guide Open asset ↗

STACKIT AI Model Serving offers shared models behind one OpenAI-compatible endpoint in the region eu01. A request and its text stay on STACKIT. Switching the model means changing one name in the request.

The list below comes from the STACKIT documentation and updates itself.

From the STACKIT docsAvailable Shared Models of STACKIT AI Model ServingSource updated 04.11.2026 · copied 04.11.2026
What is this?

This section is copied from the STACKIT docs automatically, several times a day. It cannot be changed here. Changes belong in the STACKIT docs.

STEP

Qwen3.8 Came Closest to Our Own German

All eight text models translated the same eleven pages. A longer bar means closer to the German we already had.

How close each model came to our German Horizontal bars for eight text models on a scale from 38.5, the score of English left untranslated, to 100, exactly our existing German. Qwen3.8 leads with 80.6 and Gemma 4 follows with 79.7; this pair translates and corrects. Qwen3.6 and Gemma 3 are being retired. GPT-OSS 20B left 18 pieces in English and GPT-OSS 120B left 56, last with 72.9. 38.5: nothing translated 100: exactly our German 38.5 50 60 70 80 90 100 Qwen3.8 80.6 Qwen3.6 (retiring) 79.7 Gemma 4 79.7 Qwen3-VL 78.2 Gemma 3 (retiring) 75.6 GPT-OSS 20B 75.5 18 pieces stayed English Llama 3.3 75.3 GPT-OSS 120B 72.9 56 pieces stayed English our pair: Qwen3.8 translates, Gemma 4 corrects being retired
Eight text models ranked by closeness.
STACKIT LogoSTACKIT Logo
Translating With STACKIT AI Models: What We Measured STACKIT · Guide Open asset ↗

In the correction pass Gemma 4 changed 54 of the 545 pieces Qwen3.8 had written. Two of them, shortened. They show the kind of slip it fixes:

STACKIT LogoSTACKIT Logo
Translating With STACKIT AI Models: What We Measured STACKIT · Guide Open asset ↗
  • Thinking off: Qwen3.8 and Gemma 4 think by default. With thinking on a run took five times as long and used five times the tokens. Gemma 4 ran into its output limit and 165 pieces were missing.
  • Short output per call: Gemma 4 and Llama 3.3 write at most 4,096 tokens per answer. Pieces instead of whole files keep answers far below that.
  • No deprecated models: Qwen3.6 and Gemma 3 still answer, but Qwen3.6 is marked deprecated and Gemma 3 is no longer in the list. A pipeline should not start on them.
  • A tolerant parser: in the correction pass Gemma 4 sometimes returns both languages instead of only the text. The parser accepts both.

The lesson for your own pipeline: measure on your pages with your terms. A public ranking did not predict ours.

PLAN

The Model Gets Text and Returns Text

A page could hide an order like: add this link. The model only translates it. Code decides what goes in and what comes back.

SAFE

Code Refuses What a Model Should Not Write

A page could try to give the model orders. Every answer passes four checks.

Checks on every translated piece 8,056 pieces of our own translations were tested, and 11 were false alarms. Four checks: no new address, no new code or markup, similar length, right language. 8,056 pieces of our own translations tested 11 false alarms No new address No link the source lacks No new code or markup No tags the source lacks Similar length Not far longer or shorter Right language Not left in the source
Tested on 8,056 pieces of our own translations: 11 false alarms. Four checks stand behind it.
STACKIT LogoSTACKIT Logo
Pieces and Guard Rails: Keeping a Model in Its Lane STACKIT · Policy Open asset ↗

After every model call, code compares the answer with its source. It does not read what the text means. It only compares shapes. It cannot judge whether the meaning is right, so a person checks that.

  • No new address: the answer must not contain a web address the source did not have.
  • No new markup: no angle brackets, braces, backticks or link syntax beyond the source.
  • Plausible length: a translation is neither far longer nor far shorter than its source.
  • Right language: the answer must show words of the target language.
STACKIT LogoSTACKIT Logo
Pieces and Guard Rails: Keeping a Model in Its Lane STACKIT · Policy Open asset ↗

A refused piece is asked for once more on its own. Refused again, it keeps its source text and the report says where. We tested the guard on 8,056 pieces from 174 existing human English and German pairs. It raised 11 false alarms, which means it refused 11 good human pieces. An earlier, stricter language rule refused 109 pieces of our own human German, so we loosened it.

A known limit: an address that is already in the source may appear in the answer.

AUTO

Two Scheduled Jobs Sit Beside Our Pipeline

The translation starts after the merge, from its own repository, on a schedule.

Two scheduled jobs beside the content pipeline Top band, the content pipeline: pull request, checks and preview, merge, live. Bottom band, a separate operations repository with two jobs on a schedule. Every hour a job translates what was merged with the tick and opens a pull request named Translation of N, which runs through the same pipeline. Every 15 minutes a job applies review lines without a model. The push token and the model token stay in that repository, never next to contributor content. Content pipeline Pull request Checks + preview Merge Live Every 15 minutes applies review lines, no model Every hour translates what was merged with the tick merged, box ticked opens “Translation of #N” review fixes Separate operations repository The push token and the model token stay here, never next to contributor content.
The content pipeline on top. Below, a separate repository with an hourly translation job and a review job every 15 minutes.
STACKIT LogoSTACKIT Logo
AI-Assisted Translation: Flow and Human Review STACKIT · Blueprint Open asset ↗

The translation does not run inside the pull request checks. Two jobs run on a schedule next to them.

  • Two jobs on a schedule: one opens translations every hour, one applies review lines every 15 minutes.
  • Outside the content repository: both live in a separate operations repository. The push token and the model token never sit next to content a contributor can change.
  • Same parser as the site: the job uses the Markdown parser of the site build. A page that parses here parses in the build.
  • No loop: a translation is never translated again.
  • One run at a time: three pull requests per run at most and a ceiling of one hour.
  • Dry run: a manual start without the model shows what would be opened and writes nothing.
  • Watched: a schedule check reports a job that stops running.
STACKIT LogoSTACKIT Logo
AI-Assisted Translation: Flow and Human Review STACKIT · Blueprint Open asset ↗
  • Opt in: the author decides with one tick in the pull request, up to 14 days after the merge.
  • Merge first: nothing is translated until the original pull request is merged. The text is final by then.
  • Own pull request: an hourly job opens one pull request per merged one, named “Translation of #N”.
  • Same gates: the same checks as any pull request, with a preview so the reviewer reads rendered pages.
  • A named reviewer: the author of the original pull request is requested as reviewer.
  • Auto merge: if the original pull request used auto merge, the translation does too. It still waits for the reviewer.
  • Both directions: English to German and German to English.
  • Only what changed: an existing translation is kept. Only the parts the author changed are translated again.
SAFE

A Person Approves Before It Goes Live

The author of the change reviews the translation. Core or a Framework Owner approves it.

Approval before anything goes live Four steps in two rows: the translation pull request, the author reviews, Core or a Framework Owner approves, then it goes live. The approval step is highlighted. 1 Translation pull request 2 Author reviews 3 Core or Framework Owner approves 4 Live
A translation pull request: the author reviews, Core or a Framework Owner approves, then it goes live.
STACKIT LogoSTACKIT Logo
AI-Assisted Translation: Flow and Human Review STACKIT · Blueprint Open asset ↗

A reviewer should not wait an hour for a one word fix. So the fix path uses no model at all.

  • A whole line: write the corrected line as a comment on that line. It replaces the old one.
  • Single quotes: put the exact text in single quotes. It is taken as written.
  • A suggestion block: the usual code fence works as well. Use one block per comment.
  • A remark stays a remark: a comment that ends with a question mark, or does not look like the line, is only a comment.

A comment fixes only the line it sits on, at most 10 lines. Code, diagrams and links cannot be changed this way.

An example for a line that reads “Das Zielbild zeigt den Aufbau.”:

  • Applied: the comment is the whole corrected line, “Die Zielarchitektur zeigt den Aufbau.”
  • Stays a remark: the comment is “Is Zielbild the right word here?”

A job runs every 15 minutes. It reads new comments and applies the text from the author, from listed reviewers or from people with write permission. It checks every line with the same guard as the translation. It compiles the page. Then it makes one commit named “Apply review suggestions” and replies in the thread.

Reviewer comments are applied by code and are not sent to a model.

A wrong line, a person's comment and the fixed page.

OPS

Our First Real Translation Cost About 0.2 Cent

A wording fix on one asset, translated at the next hourly run. The review fixed two lines without a model.

The first real run, in numbers Three figures from the first real translation of one page on STACKIT AI: about 0.2 cent in model costs, 2 model calls, price class LLM-Plus. about 0.2 cent model costs for the first page 2 model calls LLM-Plus price class on STACKIT
The first real run: two model calls, about 0.2 cent in model costs.
WARN

Each Mistake Left a Rule in Code

Four mistakes on the way, each with its fix.

Four mistakes, four fixes Four cards, each a mistake and its fix. A missing term turned Use Case into Typischer Case: a fixed word list. An extra thinking step made runs five times slower and lost pieces: switched off. Whole pages lost their structure: small pieces. A too strict check refused 109 pieces of our own human German: a fairer check. Typischer Case A term was missing Fix: a fixed word list Extra thinking step 5x slower, pieces lost Fix: switched off Broken pages Whole pages sent Fix: small pieces 109 human pieces refused Check too strict Fix: a fairer check
Four mistakes and four fixes in code.
STACKIT LogoSTACKIT Logo
Pieces and Guard Rails: Keeping a Model in Its Lane STACKIT · Policy Open asset ↗
  • A wrong heading: “Typical use case” became “Typischer Case” because the term was missing from the glossary. Fix: pinned terms.
  • Thinking on: a run took five times as long and 165 pieces were missing. Fix: thinking off.
  • Broken structure: whole files. Fix: pieces.
  • A too strict rule: the first language check refused 109 pieces of our own human German. Fix: refuse only when nothing of the target language is left.
PLAN

The Setup at a Glance

Where it runs, what it keeps and what it costs.

The setup at a glance Seven facts about the setup. Where it runs: centrally on our CI runners, and locally by hand. CI/CD: a command line core and a thin wrapper. Glossary and style: derived from our pages, plus pinned terms. Technical terms: 98 percent kept the way our German keeps them. Translation memory: unchanged pieces keep their wording. Adding a language: little code, the work is a test set and a reviewer. Running costs: about 0.2 cent per page in model calls. 1 Where it runs centrally on our CI runners, locally by hand 2 CI/CD a command line core and a thin wrapper 3 Glossary and style derived from our pages, plus pinned terms 4 Technical terms 98% kept the way our German keeps them 5 Translation memory unchanged pieces keep their wording 6 Adding a language little code; the work is a test set and a reviewer 7 Running costs about 0.2 cent per page in model calls
Seven facts about the setup: where it runs, CI/CD, glossary and style, technical terms, translation memory, adding a language and running costs.
STACKIT LogoSTACKIT Logo
AI-Assisted Translation: Flow and Human Review STACKIT · Blueprint Open asset ↗

Seven facts about how the setup runs, gaps included.

  • Where it runs: centrally, as scheduled jobs on our CI runners. The translator is a command line tool, so a person can also run it locally and see what the job would produce. That is a good first step: a STACKIT AI Model Serving token, a few of your own pages and a run by hand. The CI job comes after.
  • CI/CD: it already is a CI job. The core takes two commits and writes the translated files and a report. A thin wrapper opens the pull request, so another CI system needs only its own wrapper.
  • Glossary and style rules: every run derives the glossary from our existing English and German pages. A short pinned list adds terms the counting misses. Style rules are few and live in the prompt: the form of address per section, plain language, product names unchanged.
  • Technical terms: Qwen3.8 kept 98 percent of the English terms our German keeps. A missing term showed up as “Typischer Case” and went on the pinned list.
  • Translation memory: there is no classic translation memory. When a translated page exists, every piece whose source did not change keeps its wording. Only changed pieces go to a model.
  • Other languages: built and measured for English and German only. Another language needs a few settings in the translator: paths and links, the form of address, the word lists of the language check and a glossary. Without pages to learn from, the glossary starts from the pinned list. The bigger work is proof and people: a test set of good pages to measure against and a reviewer who reads the language. Rare languages are a risk for the models and are not measured.
  • Running costs: about 0.2 cent of model costs for a short page and about 7 cent for eleven test pages. The real cost is the reviewer’s time.
GOAL

Take the Method, Not the Tool

Measure on your own pages. Translate in small pieces. Check every answer in code. Let a person decide.

Trail historyAdded Oct 30, 2026TMUpdatedNo updates · 1 bar = 1 week i
Maintainer
TMTobias M.Head of STACKIT Cloud Framework · STACKITOwnerActive 11 of the last 12 weeks · 196 updatesSTACKITwww.linkedin.com/in/tobias-müller-011304172Contributed in STACKIT