Getting a decade of business logic out of the UI
Our analytics UI held ten years of business logic: how a metric is really calculated, which rules apply to which customer. An AI agent cannot reach logic welded into a UI, so it reimplements it badly and then confidently gives a different answer than the product does.
I steered and architected the work to distil that logic into a canonical query layer, shipped in increments rather than as a rewrite. We went canonical-only, with no fallback to the old path, so every gap failed loudly and set the next increment instead of quietly answering differently. The UI is one consumer of the layer, alongside our API and our agentic AI, which runs existing dashboard cards directly as steps in its workflows. Same definitions, same answer, whoever is asking.
Before a three-week break, I handed the team an initiative's worth of tickets, one document with the architecture and plan, and a chat bot seeded with the plan and my reasoning behind it, so they could answer their own questions while I was out. The step that moved dashboard cards onto the layer shipped while I was away, the headline of my catch-up notes when I got back. The team didn't need me in the room, and that's the part I'm proudest of.