An async localization job can run through an optional pipeline of stages that wrap the core translate step - AI source cleanup, human review, AI post-edit, and a back-translation drift check. Each stage is independently toggled on the localization engine, runs only for jobs created via the async API, and appears as its own step record in the job's status response.
Async API only
Pipeline stages only apply to jobs created through the Async Localization API. The synchronous /localize endpoint runs the core translate step only - pipeline settings on the engine are ignored.
On this page
- Stages at a glance
- Pre-localization AI edit
- Post-localization human review
- Post-localization AI review
- Back-translation check
- Configuring the pipeline
- Observability
Stages at a glance#
The pipeline wraps the core localization step. Any combination of stages can be enabled - disabled stages are skipped entirely, and a failure in any non-critical stage does not fail the job.
Pre-localization AI edit
Optional. An AI agent cleans the source payload - fixing typos, grammar, and spelling errors - before translation. Non-critical: if it fails, the original source is used.
Core localization
Always runs. Your localization engine applies its model config, glossary, brand voice, and instructions to produce the translation.
Post-localization human review
Optional. The translation is reviewed by a human - your own team in the dashboard's review queue (Internal Review) or a professional translator from an external provider (External Review). The workflow waits for their output before continuing.
Post-localization AI review
Optional, requires human review. An AI agent reviews the human output and adjusts it to adhere to your engine configuration.
Back-translation check
Optional. The final translation is translated back into the source locale, compared against the original, and adjusted if semantic drift is detected.
Pre-localization AI edit#
An AI agent reviews the source payload before translation, correcting typos, grammar mistakes, and spelling errors. Cleaner source text produces more consistent translations - especially across many target locales, where a single error in the source propagates to every output.
This stage is non-critical. If the pre-edit call fails or times out, the original source is passed through unchanged and the job continues.
When to enable: source content is user-generated, scraped, or otherwise likely to contain noise. Skip it for curated content that has already passed editorial review.
Post-localization human review#
The AI translation is reviewed and edited by a human for clarity, consistency, and accuracy while adhering to your engine configuration. The async job pauses until the human output is returned, then carries that output forward to any remaining stages.
Two review modes are available, selected per engine:
Internal Review#
Your own team reviews translations directly in the Lingo.dev dashboard. Pending reviews land on the Human Reviewer page of your organization (/orgs/<org-id>/human-reviewer), and reviewers are notified when new items arrive. A reviewer claims an item - claims are exclusive, so two reviewers never edit the same translation - then approves the translation as-is or submits an edited version. The job resumes immediately with their output.
Internal Review is the default mode for new engines. Use it when your team has the language expertise in-house and you want full control over the final output.
Permissions#
The Human Reviewer page is permission-gated. An organization admin grants access through roles (Settings → Roles):
| Permission | What it unlocks |
|---|---|
Review translations (engine:review_translations) | See and work the review queue - claim a pending translation, edit it, approve it as-is, or submit the edited version |
Manage reviews (org:manage_reviews) | Review history and reviewer statistics for all internal reviews across the organization |
Reviewers only need Review translations. Manage reviews is for whoever oversees the review operation - it adds the history and stats view but does not include queue access by itself.
External Review#
The translation is submitted to a qualified professional translator through an external provider. The workflow resumes when the provider returns the edited translation.
Two tiers are available for External Review:
| Tier | Best for |
|---|---|
| Standard | Accurate translation with a human voice - marketing copy, UI strings, help content |
| Pro | Professional use with an even higher accuracy level - legal, medical, and regulated content |
Timeout#
You configure how long the workflow waits for the human output - the same setting applies to both review modes. If the timeout expires with no response, the stage is marked skipped and the job continues with the AI translation as the final output. Default: 48 hours.
Event-driven wait
The workflow resumes on an event - a reviewer submitting in the dashboard (Internal Review) or a webhook from the translation provider (External Review). It is not polling on a tight loop, so setting a long timeout does not consume compute in the background.
Post-localization AI review#
After the human reviewer returns their edit, an AI agent reviews the human output and adjusts it to align with your engine's configuration - glossary terms, brand voice, and instructions. Human translators sometimes introduce variations that conflict with engine-level rules; this stage reconciles the two.
This stage is only available when post-localization human review is enabled, and only runs when the human stage actually produces output. If the human stage times out or is skipped, post-edit is also skipped — even when postEdit.enabled is true.
Not the same as AI Reviewers
Post-localization AI review modifies the translation output to conform with your engine rules. AI Reviewers are a separate feature that scores translation quality asynchronously without altering the output. You can use both together - post-edit refines the text, reviewers evaluate the result.
Back-translation check#
A multi-step quality check that verifies the translation preserves the meaning of the source.
Reverse translate
The translated output is translated back into the source locale using the same engine configuration, adapted for the reverse direction.
Detect drift
An AI agent compares the back-translation against the original source and flags issues by severity - minor, major, or critical.
Correct if needed
If major or critical issues are detected, another AI pass adjusts the forward translation to resolve them. Minor issues are recorded for observability but not auto-corrected.
Back-translation is a classic QA technique in human translation workflows. The pipeline automates it using LLMs, surfacing semantic drift that would otherwise require a second pair of human eyes.
Configuring the pipeline#
Pipeline configuration has two layers: an engine-level default and an optional per-request override.
Engine-level configuration#
Open the engine's Pipeline tab in the dashboard and toggle each stage independently. The engine's configuration applies to every async job routed to it when no override is provided.
- Post-localization human review runs through Internal Review (your team, in the dashboard) or External Review (professional translators) - pick the mode, tier, and timeout in the same panel.
- Post-localization AI review is disabled until human review is enabled.
- Back-translation check is independent of the other stages.
Per-request override#
Pass a pipelineConfig object in the POST /jobs/localization body to override stages for a single submission. Stages you omit inherit from the engine config.
{
"sourceLocale": "en",
"targetLocales": ["de", "fr"],
"data": { "title": "Hello" },
"pipelineConfig": {
"preEdit": { "enabled": true },
"backTranslation": { "enabled": false }
}
}Use this for one-off submissions that need a different pipeline than the engine's default, without creating a new engine.
Observability#
Every enabled stage produces a step record on the localization job. Retrieve the job with GET /jobs/localization/:jobId and inspect the steps array:
{
"id": "ljb_...",
"status": "completed",
"outputData": { "title": "Hallo" },
"steps": [
{
"stepId": "preEdit",
"type": "action",
"status": "completed",
"errorMessage": null,
"createdAt": "2026-04-17T10:00:00Z",
"startedAt": "2026-04-17T10:00:01Z",
"completedAt": "2026-04-17T10:00:04Z"
},
{
"stepId": "localize",
"type": "action",
"status": "completed",
"errorMessage": null,
"createdAt": "2026-04-17T10:00:04Z",
"startedAt": "2026-04-17T10:00:04Z",
"completedAt": "2026-04-17T10:00:12Z"
}
]
}Each step's stepId identifies the stage:
| Stage | stepId |
|---|---|
| Pre-localization AI edit | preEdit |
| Core localization | localize |
| Post-localization human review | humanEdit |
| Post-localization AI review | postEdit |
| Back-translation check | backTranslation |
Each step is marked completed, failed, or skipped independently. Non-critical failures do not fail the job - the last successful output carries forward to the final outputData, and the job status becomes completed_with_warnings with per-step messages in the job's warnings array. See Reports for aggregate stage success rates.