Baemin's 5-Year Multilingual Homework, Finished in a Month with AI

Baemin attempted the project for 5 years to support hundreds of thousands of stores and tens of millions of menus in multiple languages, but failed every time due to the massive data scale and the difficulty of establishing context-aware translation standards and operations.

In early 2026, a breakthrough came as the FDH (Food Data Hub) system for processing large-scale data converged with improvements in LLM (Large Language Model) quality.

The implemented architecture largely consists of two pipelines.

• FDH Translation Pipeline: When menu data changes, an event is published, and the LLM performs translation asynchronously, loading the results as metadata.
• Backend BFF Handling: It analyzes the Accept-Language header of app requests, maps the loaded multilingual data, and returns a response.

This approach allows multilingual support to be applied quickly through server deployment alone, without an app update, and enables gradual quality improvement.