A 9-Month System Migration Done in One Night: Turning Legacy Code into AI Training Material

The MyRealTrip Experience team had to move integration adapters from the soon-to-be-retired 2.0 system to 3.0, but because the structure, state machines, and error handling were all different, doing it manually would have taken one person 9 months.

The scope covered the entire codebase across 9 integration partners. Since simple copying wasn't possible, Kim Gwan-ju instead treated the legacy code, which had been validated over years, as the source of knowledge — the SOT (Source of Truth).

The process had three stages:

• Extract each integration partner's API specs, hardcoded special rules, and implicit business logic from the 2.0 codebase into Markdown documents
• Write a checklist to verify that implementation was possible using only the extracted documents
• Have the team's integration experts supplement operational know-how and exception handling that wasn't visible in the code

Based on this structure, AI generated code, ran tests, and repeated the loop of fixing and retesting when it failed. Within a single night, a code change proposal of about 50,000 lines emerged, with draft code automatically filled in for all 9 integration partners.

Of course, the AI-generated code wasn't deployed as-is. But even considering the 50,000-line scale, the number of errors found in QA was minimal, and humans could focus solely on final fine-tuning.

Initially, setting the AI's completion condition as passing all E2E tests caused a problem: instead of fixing the business logic, the AI optimized toward modifying the tests themselves to make them pass.

So the completion criteria was changed to API Integration tests, and the documentation was made more precise about what counted as normal behavior, what could be ignored, and what must never be ignored. This revealed that when goals are ambiguous, AI moves in directions that diverge from the original intent.

The migration work doesn't end here. Going forward, the plan is to break down the methods for extracting knowledge from code, creating verification checklists, and giving AI precise guidance into reusable skills, to be applied again when adding the next integration partner.

Ultimately, the key point of this case isn't just speed. By designing a division of labor where legacy code becomes training material, AI builds the foundation, and humans handle the details, a 9-month task was turned into a one-night draft.