---
title: "Translating in Our Pipeline With STACKIT AI Models"
description: "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."
scfTrail:
  tags: ["translation", "STACKIT AI", "Model Serving", "pipeline", "human in the loop"]
  maintainers:
    - user: "tobias.mueller"
  steps:
    - style: "compass"
      id: "problem"
      title: "A Draft in Seconds Is Not Yet a Translation"
      trailContext: "A model writes fast. It can also break a link or invent a sentence."
      imageSrc: "contributors/stackit/files/data-and-ai/translation-gap.svg"
      imageAlt: "About 140 English assets without German, 6 minutes for eleven test pages, one person who checks."

    - role: sub
      id: "problem-detail"
      title: "The gap in detail"
      trailContext: "Why a fast draft alone is not enough."
      assetId: "data-and-ai/assetcontainer/stackit/translation-pipeline-flow#the-problem"

    - style: "gondola"
      id: "route"
      title: "One Tick Starts the Whole Route"
      trailContext: "The author decides. STACKIT AI translates once the change is live. The author reviews the result."
      imageSrc: "contributors/stackit/files/data-and-ai/translation-flow.svg"
      imageAlt: "Five stages from one tick to an approved translation."

    - style: "compass"
      id: "tests"
      title: "Five Tests Before the First Real Run"
      trailContext: "We tested on our own pages before a single real pull request was translated."
      imageSrc: "contributors/stackit/files/data-and-ai/translation-tests.svg"
      imageAlt: "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."

    - style: "stairs"
      id: "pieces"
      title: "Small Pieces Keep Pages Intact"
      trailContext: "Sent whole, 2 to 4 of 11 pages broke. In small pieces our chosen model kept all 11 intact."
      imageSrc: "contributors/stackit/files/data-and-ai/translation-placeholders.svg"
      imageAlt: "A command becomes a marker and comes back after the translation."

    - role: sub
      id: "pieces-travel"
      title: "How a piece travels"
      trailContext: "Split, protect, translate, correct, restore."
      assetId: "data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails#how-a-piece-travels"

    - style: "hut"
      id: "models"
      title: "Open Models Run on STACKIT AI Model Serving"
      trailContext: "Switching a model means changing one name. The text stays on STACKIT."
      imageSrc: "contributors/stackit/files/data-and-ai/translation-model-lineup.svg"
      imageAlt: "Qwen3.8 translates and Gemma 4 corrects. More open models run on the same endpoint."

    - role: sub
      id: "models-list"
      title: "The model list"
      trailContext: "Type, context, output limit and status of each model."
      assetId: "data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models#the-models-on-stackit-ai-model-serving"

    - style: "stairs"
      id: "measure"
      title: "Qwen3.8 Came Closest to Our Own German"
      trailContext: "All eight text models translated the same eleven pages. A longer bar means closer to the German we already had."
      imageSrc: "contributors/stackit/files/data-and-ai/translation-models.svg"
      imageAlt: "Eight text models ranked by closeness."

    - role: sub
      id: "measure-correct"
      title: "A second model corrects"
      trailContext: "Gemma 4 changed 54 of 545 pieces Qwen3.8 had written, for example a wrong article."
      assetId: "data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models#a-second-model-corrects"

    - role: sub
      id: "measure-settings"
      title: "Settings that mattered"
      trailContext: "Thinking off, short answers, no deprecated models."
      assetId: "data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models#settings-that-mattered"

    - style: "compass"
      id: "principle"
      title: "The Model Gets Text and Returns Text"
      trailContext: "Every later step keeps this rule true."
      center:
        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."

    - style: "shield"
      id: "guard"
      title: "Code Refuses What a Model Should Not Write"
      trailContext: "A page could try to give the model orders. Every answer passes four checks."
      imageSrc: "contributors/stackit/files/data-and-ai/translation-guard.svg"
      imageAlt: "Tested on 8,056 pieces of our own translations: 11 false alarms. Four checks stand behind it."

    - role: sub
      id: "guard-checks"
      title: "The four checks"
      trailContext: "What the guard compares."
      assetId: "data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails#what-the-guard-checks"

    - role: sub
      id: "guard-refuses"
      title: "When the guard refuses"
      trailContext: "Asked once more, then the source text stays and the report says where."
      assetId: "data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails#when-the-guard-refuses"

    - style: "gondola"
      id: "pipeline"
      title: "Two Scheduled Jobs Sit Beside Our Pipeline"
      trailContext: "The translation starts after the merge, from its own repository, on a schedule."
      imageSrc: "contributors/stackit/files/data-and-ai/translation-pipeline.svg"
      imageAlt: "The content pipeline on top. Below, a separate repository with an hourly translation job and a review job every 15 minutes."

    - role: sub
      id: "pipeline-detail"
      title: "How it fits into the pipeline"
      trailContext: "Two jobs, one parser, no loop, a dry run."
      assetId: "data-and-ai/assetcontainer/stackit/translation-pipeline-flow#how-it-fits-into-our-pipeline"

    - role: sub
      id: "route-detail"
      title: "Each stage of the flow"
      trailContext: "Opt in, merge first, own pull request, same gates, named reviewer."
      assetId: "data-and-ai/assetcontainer/stackit/translation-pipeline-flow#the-flow"

    - style: "shield"
      id: "human"
      title: "A Person Approves Before It Goes Live"
      trailContext: "The author of the change reviews the translation. Core or a Framework Owner approves it."
      imageSrc: "contributors/stackit/files/data-and-ai/translation-approval.svg"
      imageAlt: "A translation pull request: the author reviews, Core or a Framework Owner approves, then it goes live."

    - role: sub
      id: "human-fix"
      title: "Fixing a line"
      trailContext: "Whole line, single quotes or a suggestion block. A remark stays a remark."
      assetId: "data-and-ai/assetcontainer/stackit/translation-pipeline-flow#fixing-without-a-model"

    - style: "chart"
      id: "cost"
      title: "Our First Real Translation Cost About 0.2 Cent"
      trailContext: "A wording fix on one asset, translated at the next hourly run. The review fixed two lines without a model."
      imageSrc: "contributors/stackit/files/data-and-ai/translation-cost.svg"
      imageAlt: "The first real run: two model calls, about 0.2 cent in model costs."

    - style: "crevasse"
      id: "lessons"
      title: "Each Mistake Left a Rule in Code"
      trailContext: "Four mistakes on the way, each with its fix."
      imageSrc: "contributors/stackit/files/data-and-ai/translation-lessons.svg"
      imageAlt: "Four mistakes and four fixes in code."

    - role: sub
      id: "lessons-detail"
      title: "The four mistakes"
      trailContext: "Each mistake with its fix."
      assetId: "data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails#what-went-wrong-on-the-way"

    - style: "compass"
      id: "setup"
      title: "The Setup at a Glance"
      trailContext: "Where it runs, what it keeps and what it costs."
      imageSrc: "contributors/stackit/files/data-and-ai/translation-setup.svg"
      imageAlt: "Seven facts about the setup: where it runs, CI/CD, glossary and style, technical terms, translation memory, adding a language and running costs."

    - role: sub
      id: "setup-detail"
      title: "The setup in detail"
      trailContext: "Including the gaps: no classic translation memory and two languages so far."
      assetId: "data-and-ai/assetcontainer/stackit/translation-pipeline-flow#the-setup-at-a-glance"

    - style: "summit"
      id: "takeaway"
      title: "Take the Method, Not the Tool"
      trailContext: "Four rules we would apply to any text a model writes for us."
      center:
        text: "Measure on your own pages. Translate in small pieces. Check every answer in code. Let a person decide."

  presentations:
    - id: "technical"
      label: "Technical deep-dive, 25 minutes"
      description: "Every step with the measurements, settings and limits behind it."
      launch: true
      preset: "focus"
      steps:
        - { id: problem, notes: "We measured a model on eleven of our own pages: 6 minutes for all of them. Speed was solved on the first day. The rest of this talk is about trust: links that must not break, terms that must stay English and a person who has the last word." }
        - { id: route, notes: "This is where we ended up. The box sits in the pull request template, nothing runs before the merge and a person approves at the end. The next slides show what we tested before we built it." }
        - { id: tests, notes: "Every one of these ran on our own pages before the first real pull request. The next slides open them one by one. The dry run read the last 150 closed pull requests: none had ticked the box yet, so it rightly opened nothing." }
        - { id: pieces, notes: "A parser cuts the page into text pieces. Links, inline code, components and image targets become numbered markers. Code blocks and imports never leave. Sent as a whole file, models translated YAML inside code examples and a badge value that must stay English. To test the cutter alone we rebuilt 222 pages without any model. All came back byte for byte." }
        - { id: pieces-travel, role: sub, notes: "German links get the de/ prefix and German anchors by code. The form of address is read off our own German pages: du in the how-tos, the getting-started trails, ISV and the contributor profiles, Sie everywhere else. The model is told which one to use." }
        - { id: models, notes: "One OpenAI-compatible endpoint in eu01, so switching a model means changing one name. Qwen3.8 writes up to 16,384 tokens per answer, Gemma 4 only 4,096. Qwen3.6 is marked deprecated and Gemma 3 is no longer in the list, so a pipeline should not start on them. The full model list in the trail comes live from the STACKIT documentation, which is why the slide names no number." }
        - { id: measure, notes: "Closeness is the chrF score: matching letter sequences against our existing German. English left as is already scores 38.5. GPT-OSS is fast but left up to 56 pieces in English. Four models kept only 10 of 11 files intact even in small pieces. The two leaders fail in different ways: Qwen3.8 slips on gender and case, Gemma 4 mixes in too many English words. That is why they work as a pair." }
        - { id: measure-correct, role: sub, notes: "Gemma 4 changed 54 of 545 pieces. The other way round almost nothing changed. Sometimes Gemma 4 answers with both languages instead of only the text, so the parser accepts both." }
        - { id: measure-settings, role: sub, notes: "Thinking on cost five times the time and the tokens. Gemma 4 then ran out of room and 165 pieces were missing. So thinking is off for both models." }
        - { id: principle, notes: "This rule came from one question: how can a contributor improve a translation without steering the AI by prompt? Our answer: comments are applied by code and never sent to a model. The open side is the page itself, because its English text goes to the model. That is why every answer passes a guard." }
        - { id: guard, notes: "The guard checks shape, not meaning. We tuned it on 174 pairs of our own human translations: 8,056 pieces and 11 false alarms. Known limit: an address that is already in the source may appear in the answer. A first, stricter language rule refused 109 pieces of our own human German, so now a piece is only refused when nothing of the target language is left." }
        - { id: guard-refuses, role: sub, notes: "A refused piece gets one more try on its own. Refused again, the source text stays and the report names the place. A German page with one English sentence is better than a page with a sentence nobody asked for." }
        - { id: pipeline, notes: "Both jobs live in a separate operations repository, not in the content repository. A workflow runs with the secrets of the repository it lives in, and a pushed branch runs the workflow as written in that branch. So the push token and the paid model token must not sit next to content. One merged pull request gives one translation pull request, at most three per hourly run and one run at a time. A translation never starts another translation. The job uses the Markdown parser of the site build, so a page that parses here parses in the build." }
        - { id: route-detail, role: sub, notes: "The author of the change is the requested reviewer. Core or a Framework Owner approves. If the original pull request used auto merge, the translation does too. It still waits for the reviewer." }
        - { id: human, notes: "The requested review holds the auto merge. A contributor review is never the official one, so Core or a Framework Owner still approves. Our first real fix was one word, and waiting an hour for it felt too long. So review suggestions now run every 15 minutes, without a model." }
        - { id: human-fix, role: sub, notes: "Nobody wants to type a suggestion fence. So a plain line counts when at least half its words match the old line and it does not end with a question mark. Single quotes make it literal. The job adds heading marks or list dashes itself. On #378 two lines were taken and one remark stayed a remark." }
        - { id: cost, notes: "The first real one was #377, a wording fix on the streaming analytics asset. Its translation #378 opened at the next hourly run. Two model calls for one short page, about 0.2 cent. The eleven test pages came to about 7 cent. Both models are in the LLM-Plus class at 0.45 and 0.65 euro per million tokens. The real cost is the reviewer's time." }
        - { id: lessons, notes: "Typical use case became Typischer Case. Our glossary was read off our own pages and only found capitalised phrases. Now nine pinned terms, like Use Case and Best Practices, always stay English. One trap we nearly fell into: German writes many terms with a hyphen, Cloud-Transformation 68 times and never with a space. Those must not go on a keep-English list." }
        - { id: setup, notes: "Two of these are gaps, said plainly: there is no classic translation memory, and only English and German are built and measured. The code needs a few settings per language. The real work for another language is a test set to measure against and a reviewer who reads it." }
        - { id: takeaway, notes: "None of this is special to translation. Any text a model writes for a public site needs the same four things." }
    - id: "overview"
      label: "Overview for decision-makers, 10 minutes"
      description: "Why we do it, how it stays safe and what it costs. No technical terms."
      steps:
        - { id: problem, notes: "Translation by hand is slow, so the German side kept falling behind. An AI translated eleven of our pages in six minutes. Speed was easy. Trust was the question." }
        - { id: route, notes: "The author decides with one tick. Nothing is translated before the change is live. The author reviews the translation and Core or a Framework Owner approves it. Nothing goes live before that." }
        - { id: tests, notes: "Before the first real page we tested five things on our own pages: whole pages against small pieces, the cutting alone, eight models, the safety checks and a dry run over old pull requests. Only then did we switch it on." }
        - { id: pieces, notes: "We cut every page into small pieces. Links and commands are swapped for markers before the AI sees the text and put back afterwards. Whole pages broke in some of our tests. In small pieces our chosen model kept every page intact." }
        - { id: measure, notes: "We only use models that run on STACKIT. The text never leaves the platform. We tested eight of them on our own pages. Qwen3.8 came closest to the German we already had. A second model then fixes small grammar slips, because the two make different mistakes." }
        - { id: principle, notes: "Whatever a page says, the AI is only asked to translate it. Comments from reviewers never reach the AI at all. If an answer adds a link, code or markup, our checks throw it away." }
        - { id: guard, notes: "This is the safety net. We tested it on 174 pages of our own translations and it wrongly refused 11 of 8,056 good pieces. Whether the meaning is right, the person checks." }
        - { id: pipeline, notes: "The translation does not slow down the normal pipeline. Two jobs run on a timer next to it: one translates every hour, one applies review fixes every 15 minutes. The keys that may write and that pay for the AI stay in a separate place, away from content anyone can change." }
        - { id: human, notes: "A person always has the last word. When we fixed our first word, waiting an hour felt too long. Now a small fix takes a comment and is applied within 15 minutes." }
        - { id: cost, notes: "About 0.2 cent per page in model costs, measured on the first real run. Eleven test pages cost about 7 cent. The main effort stays with the person who checks the result." }
        - { id: setup, notes: "In short: it runs centrally, another CI system needs only a small wrapper and a page costs a fraction of a cent. A new language is mostly a test set and a reviewer who reads it." }
        - { id: takeaway, notes: "These four rules are the method. They apply to any text an AI writes for us, not only to translations." }
source_url: "https://framework.stackit.cloud/data-and-ai/trails/stackit/ai-assisted-translation/"
source_file: "docs/data-and-ai/trails/stackit/ai-assisted-translation.mdx"
---

## Steps

### 1. A Draft in Seconds Is Not Yet a Translation

Stage: `compass`

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

### 2. The gap in detail

Why a fast draft alone is not enough.

Asset: [/data-and-ai/assetcontainer/stackit/translation-pipeline-flow/#the-problem](/data-and-ai/assetcontainer/stackit/translation-pipeline-flow/#the-problem) — source: [/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-flow.md](/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-flow.md), section `#the-problem`

### 3. One Tick Starts the Whole Route

Stage: `gondola`

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

### 4. Five Tests Before the First Real Run

Stage: `compass`

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

### 5. Small Pieces Keep Pages Intact

Stage: `stairs`

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

### 6. How a piece travels

Split, protect, translate, correct, restore.

Asset: [/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails/#how-a-piece-travels](/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails/#how-a-piece-travels) — source: [/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails.md](/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails.md), section `#how-a-piece-travels`

### 7. Open Models Run on STACKIT AI Model Serving

Stage: `hut`

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

### 8. The model list

Type, context, output limit and status of each model.

Asset: [/data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models/#the-models-on-stackit-ai-model-serving](/data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models/#the-models-on-stackit-ai-model-serving) — source: [/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models.md](/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models.md), section `#the-models-on-stackit-ai-model-serving`

### 9. Qwen3.8 Came Closest to Our Own German

Stage: `stairs`

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

### 10. A second model corrects

Gemma 4 changed 54 of 545 pieces Qwen3.8 had written, for example a wrong article.

Asset: [/data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models/#a-second-model-corrects](/data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models/#a-second-model-corrects) — source: [/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models.md](/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models.md), section `#a-second-model-corrects`

### 11. Settings that mattered

Thinking off, short answers, no deprecated models.

Asset: [/data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models/#settings-that-mattered](/data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models/#settings-that-mattered) — source: [/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models.md](/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-stackit-models.md), section `#settings-that-mattered`

### 12. The Model Gets Text and Returns Text

Stage: `compass`

Every later step keeps this rule true.

### 13. Code Refuses What a Model Should Not Write

Stage: `shield`

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

### 14. The four checks

What the guard compares.

Asset: [/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails/#what-the-guard-checks](/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails/#what-the-guard-checks) — source: [/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails.md](/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails.md), section `#what-the-guard-checks`

### 15. When the guard refuses

Asked once more, then the source text stays and the report says where.

Asset: [/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails/#when-the-guard-refuses](/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails/#when-the-guard-refuses) — source: [/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails.md](/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails.md), section `#when-the-guard-refuses`

### 16. Two Scheduled Jobs Sit Beside Our Pipeline

Stage: `gondola`

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

### 17. How it fits into the pipeline

Two jobs, one parser, no loop, a dry run.

Asset: [/data-and-ai/assetcontainer/stackit/translation-pipeline-flow/#how-it-fits-into-our-pipeline](/data-and-ai/assetcontainer/stackit/translation-pipeline-flow/#how-it-fits-into-our-pipeline) — source: [/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-flow.md](/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-flow.md), section `#how-it-fits-into-our-pipeline`

### 18. Each stage of the flow

Opt in, merge first, own pull request, same gates, named reviewer.

Asset: [/data-and-ai/assetcontainer/stackit/translation-pipeline-flow/#the-flow](/data-and-ai/assetcontainer/stackit/translation-pipeline-flow/#the-flow) — source: [/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-flow.md](/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-flow.md), section `#the-flow`

### 19. A Person Approves Before It Goes Live

Stage: `shield`

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

### 20. Fixing a line

Whole line, single quotes or a suggestion block. A remark stays a remark.

Asset: [/data-and-ai/assetcontainer/stackit/translation-pipeline-flow/#fixing-without-a-model](/data-and-ai/assetcontainer/stackit/translation-pipeline-flow/#fixing-without-a-model) — source: [/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-flow.md](/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-flow.md), section `#fixing-without-a-model`

### 21. Our First Real Translation Cost About 0.2 Cent

Stage: `chart`

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

### 22. Each Mistake Left a Rule in Code

Stage: `crevasse`

Four mistakes on the way, each with its fix.

### 23. The four mistakes

Each mistake with its fix.

Asset: [/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails/#what-went-wrong-on-the-way](/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails/#what-went-wrong-on-the-way) — source: [/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails.md](/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-guard-rails.md), section `#what-went-wrong-on-the-way`

### 24. The Setup at a Glance

Stage: `compass`

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

### 25. The setup in detail

Including the gaps: no classic translation memory and two languages so far.

Asset: [/data-and-ai/assetcontainer/stackit/translation-pipeline-flow/#the-setup-at-a-glance](/data-and-ai/assetcontainer/stackit/translation-pipeline-flow/#the-setup-at-a-glance) — source: [/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-flow.md](/raw/data-and-ai/assetcontainer/stackit/translation-pipeline-flow.md), section `#the-setup-at-a-glance`

### 26. Take the Method, Not the Tool

Stage: `summit`

Four rules we would apply to any text a model writes for us.

