Production proof · Origin system - The system that taught me AI workflows have to be operated, not just built.
Co-op Translator and Localizeflow turn changing source content into synchronized multilingual pull requests across Microsoft open-source learning repositories.
- System
- Co-op Translator + Localizeflow
- Role
- Creator · Maintainer · Operator
- Environment
- Microsoft OSS · GitHub · Azure
- Status
- Public · Operating
Simplified production flow
A changing document becomes a traceable multilingual update.
- 01
Source change
GitHub repository
- 02
Detect and plan
Localizeflow
- 03
Translate and validate
Co-op Translator
- 04
Run at scale
Azure Container Apps Jobs
- 05
Open updates
GitHub pull requests
What this system is
Two layers, one operating workflow.
Co-op Translator is the open-source AI processing engine. It translates Markdown, notebooks, and images with Azure AI services and packages the work into a GitHub-friendly workflow.
Localizeflow is the production layer built as the workload grew beyond a single CI job. It coordinates distributed execution on Azure, tracks state, retries failed work, and returns changes through pull requests that people can inspect.
- Azure OpenAI
- Azure AI Vision
- Azure Container Apps Jobs
- GitHub App
- Pull requests
The operating problem
Translation was not a one-time task. It was a system that had to keep up with change.
The first goal was simple: make a Microsoft learning repository available in Korean and keep it current as the original changed.
At repository scale, the difficult work moved beyond translation quality. The system had to find changes, split large workloads, preserve document structure, recover from failures, and deliver reviewable updates without losing track of state.
How the pieces fit
The engine stays open. The production layer makes it operable at scale.
Co-op Translator
Open-source translation engine and workflow. It handles content parsing, AI processing, structure preservation, and translation output.
Localizeflow
Azure-hosted orchestration and product layer. It handles repository connection, distributed jobs, run state, retries, and pull-request delivery.
- Microsoft learning repositories in the current operating record
- 17
- languages in the current translation workflow
- 55
- translation tasks operated through Localizeflow
- 8,000+
- distributed execution and production hosting
- Azure
These are current operator-maintained figures. Public pages were published at different points in the project and may show earlier adoption snapshots.
What this proves
- A complex AI workflow can be decomposed into observable, recoverable stages.
- Azure-based execution can support workloads that outgrow a single CI job.
- Human review can remain in the loop through inspectable pull requests.
- The same workflow can be maintained across many repositories and languages.
What this does not claim
- It is not a client delivery completed by Reliable Workflow.
- It is not proof that every business process should be automated.
- It does not remove the need for owners, permissions, review, or exception handling.
- It is not presented as an official Microsoft product.
Public evidence
Follow the work, not a sales claim.
The engine, Microsoft community reports, and the production layer are publicly inspectable. Together they show how the system developed and how it is used.