A year building an AI design system taught me the machine could match the surface and never the system underneath. The harder problem was the one it left behind: when a senior person leaves, the judgment they carry leaves with them.
Role
Creator and primary researcher
Organization
Oracle Health
This project started as a defense to a hypothetical situation: A PM uses Figma Make to create a product workflow that looks like Oracle Redwood (but isn’t) and ships it to engineering, without design ever seeing it.
The best way to stop off-system work would be to make the system available to everyone. If a design system’s rationale was machine accessible, then a PM generating a screen would be working from the same thing a designer works from.
I worked toward designing a knowledge layer that lived outside specific design tools, so anything generating a screen could query the design rationale instead of guessing at it.
The architecture I designed for Oracle isn’t the one I’d prescribe for anyone else. In fact, I don’t think there is a one-size-fits-all solution to this particular problem.
Then the layoff came, and slowing down was what let me see straight. We’d been moving so fast that none of us caught the holes in our own thinking. We’d been satisfied the internal documentation retrieval was the destination, when it was only the surface.
When any person leaves, their institutional knowledge leaves with them. How can AI learn those non-documented nuggets? Especially when that information fixes problems in the documented knowledge?
The knowledge in layers, each owned by the people who produce it.
The plain version of this idea is that you capture what’s in people’s heads before they leave, before they quit or get walked out. Said out loud, it lands somewhere between pragmatic and dystopian, and I’ve never fully made peace with which. The benefit of better context is real, but I can’t pretend the other edge isn’t there. Forcing what’s in someone’s head out into the open, so a company keeps it after the person is gone, is useful but far from settled. The bigger parts of this need to be answered by laws and labor norms and voting, not by designers and engineers.
We shipped three components on Oracle’s Redwood system. The two I remember best, we killed: one would have outed a patient, one broke the rule that keeps a name reliable.
I had twelve reports across six locations and three continents, through a $28 billion acquisition and two years with no promotions on the table. Three of them got promoted anyway.
Scouting America has a retention problem and a volunteer position meant to fix it that almost nobody fills. A proposal for an AI coordinator that does the work the volunteers can’t keep up with.
