Medical
diagnostic copilot

The real risk isn't that the machine gets it wrong.

Medical AI always sells the same promise: « it finds the right diagnosis ». That's the wrong problem. In a clinical setting, the failure to fear is automation bias — the moment a competent practitioner sees a confident display, stops exercising critical judgment, and follows. In medicine, that failure can kill.

So I inverted the objective. Not « what is the best answer? » but « how do you keep the physician reasoning while speeding up their work? » It's a cognitive-ergonomics problem before it's an AI problem — the intersection I've worked at for eighteen years.

End-to-end design direction and prototyping: problem framing, interaction principles, information architecture, three working screens. Prototypes built by directing agentic generation tools — I design, the tool executes. The approach itself is a demonstration of an AI-native workflow: designing agent interfaces by working with agents.

Structuring the interface so it forces reasoning instead of short-circuiting it.

A table where every hypothesis is readable at a glance. For each, the signs already gathered are set apart from the signs still to explore — the display becomes an interview aid. Signs are added on the fly, including through active listening: while the patient speaks, the differential re-weights in real time. Connected patient context, a « focus » panel per hypothesis, and a closing that writes to the record.

consultation.differential · live

The view that adds time. The differential has a past. The tree replays how the reasoning evolved entry by entry, and the agent flags the pivots — the moments the picture swung hard. Every node is traceable: gathered elements, sources, and for a dismissed hypothesis, the reason for dismissal. You can go back to an earlier state: if a lead turns out wrong, the physician restarts from the moment before. Clinical accountability stays documented and reversible.

reasoning.tree · historised

A good concept owns its grey areas.

• Link between the two views. Tabs? The tree as a « zoom-out » of the consultation? A trade-off between cognitive continuity and clarity.
• Archive vs branch on going back — keep one path, or both.
• The risk of « to-explore » signs. They say what to ask, but could turn the consultation into a checklist and reduce reasoning. Sketched answer: let the physician add their own leads, so the human extends the machine's frame rather than submitting to it.