Banksalad Automatically Generates MyData Test Data Using LLM

Due to the nature of MyData services, generating test data that reflects complex financial product information and user patterns was a major challenge for Banksalad's QA team, but this was solved by adopting LLM. By learning from real data, LLM can generate realistic data reflecting consumption patterns, seasonal variations, and socioeconomic factors.

Prompt Engineering and Data Processing

By setting specific personas and specifying JSON format, they guided the LLM to understand context and generate appropriate data. The generated data is structured through post-processing steps including markdown removal, deletion of unnecessary explanations, JSON parsing, and validation.

Automation and Utilization

They automated the process of mapping structured JSON data to Kotlin/Java DataClass and sending it to the development server via API. Using GitHub Actions' Cron Job, new data is generated daily, and this is shared not only for QA testing but also in the development environment, improving efficiency in testing features that require long-term history.