A model response is never just one state. It streams, it stalls, it comes back uncertain, wrong, or empty — and every one of those is an interface someone has to read. I design each of them on purpose, tokenise them, and document them so engineers and data scientists build against the same picture.

The work starts before a prompt exists: short, structured discovery workshops with the people who own the problem, leaving with the two or three AI-supported scenarios worth building, the risks written down, and an agreed definition of done. From there it is user journeys, information architecture, wireframes and high-fidelity design in Figma, on an atomic component structure and a documented design system — so the foundations carry between features instead of being redrawn for each one.

Accessibility is part of the interaction model rather than a final polish pass. Contrast is measured against the tokens actually shipping, every control is keyboard-operable, and state changes are announced rather than only painted.