Designing Around the Model
I joined NYU Langone's HiBRID Lab in June 2026. In the months since, I've worked across four projects that share nothing on the surface — a chatbot proposal for cardiology, a decision-support tool for statins, a behavioral-nudge system for primary care, a content pipeline for the lab itself.
What actually connected them wasn't the clinical subject matter. It was three questions nobody wanted to spend real design time on: what can this model do, what can it not do, and how long does it take to find out? Every interface decision I made traced back to one of those three — which is what this case study is actually about. It's still in motion; so is the thinking.
Clinical AI gets graded on whether it's right. In practice, that was rarely what decided whether it actually worked. A correct recommendation that arrives after the patient has already left, or one hedged behind data nobody knew was missing, changes nothing.
How might we design clinical AI around what the model can't do, rather than around what it can?
- Designed a statin decision-support tool alongside a physician lead and a second UX designer
- Ran and coded usability sessions with prescribers, with a second observer for reliability
- Built a journey-mapping tool coordinators use to flag exactly where information disappears
- Built an automated content pipeline on the Anthropic API — intake form to review queue, no manual drafting
- Turned research findings into placement and content decisions across four Epic surfaces
So instead of pitching a chatbot, I built a journey tool. Coordinators used it to mark exactly where information disappears across eight phases of care. The capability question got answered with evidence instead of enthusiasm — and the proposal is still waiting on what it found.
The pipeline drafts captions and fill sheets for the lab from one intake form. It could have been an open text box. Instead it runs on eight fixed formats, each with its own layout and prompt — constrained enough to hand to a reviewer instead of a rewriter.
Model calls are slow enough that a synchronous "generate" button would've felt broken, so it runs on background functions with an approval queue instead. Nobody waits on the model.
The tool automates risk stratification across a risk score, guideline enhancers, and imaging signals. What mattered most wasn't the recommendation — it was the negative space: which biomarkers were never measured, which imaging was never ordered. So it states its recommendation, attaches the guideline class, and weights Order and Decline equally. It never places the order itself.
In usability sessions, prescribers noticed nudges, agreed with them, and moved on anyway — past the moment it could matter. The gap between agreeing and acting was almost entirely timing. It got sharper in the data: a behavioral health score takes about fifteen minutes to finish computing after a visit ends. The patient's already gone. A number that could have been a conversation becomes a PDF instead.
Three months in, most of this is still moving. The content pipeline is live — the lab is using it to draft real posts, not test data. The statin tool is in usability testing with prescribers, with a phased pilot planned for later this year. The cardiology chatbot is still a proposal, waiting on the capability question the journey tool was built to answer. The nudge research is feeding directly into how the behavioral health score gets redesigned next.
That's not a tidy ending. It's a status update from inside a project that isn't finished — which is the honest way to write about work you're still in the middle of.