Building an LLM-assisted refactoring workflow for a banking API reference

The situation

Mambu documents its platform in a large suite of API references. Descriptions written by many people over many releases tend to drift apart, and inconsistencies and inaccuracies are hard to spot by reading page by page, let alone fix across a whole suite by hand.

What I did

• Designed an LLM-assisted refactoring program for the full API reference suite, so the work could run reference by reference to one standard instead of as a one-off cleanup.
• Built quality-measurement tooling to find inconsistent, unclear, or inaccurate descriptions and to show whether each rewrite actually improved them.
• Rewrote the first API references with LLM assistance, under editorial review, bringing descriptions into a consistent style and correcting technical inaccuracies.

Result

The first API references, about 20% of the suite, were completed before my time at Mambu ended, with clear gains in consistency, clarity, and accuracy. Just as important, the workflow and tooling now exist to take on a reference suite of that size.