AI-powered goal setting for chronic condition patients — a product problem I defined, prioritized, and drove to completion as design lead, product manager, and data scientist.
Codex's business model required scaling clinicians from managing roughly 50 patients to 1,000 or more. Every minute a clinician spends on something the product could handle directly limits how many patients they can serve. Design wasn't a support function here — it was how the scaling actually happened.
I joined without a dedicated product manager. I owned problem definition, roadmap prioritization, design, and delivery across the patient app and clinician platform.
I established weekly research meetings with the clinical team — they didn't exist before I joined — and used them to build a prioritized roadmap across the full product surface: education, scheduling, clinical triaging, visit quality, communication tools. Goal-setting kept rising to the top, for three converging reasons.
This wasn't a product roadmap handed to me — it was a prioritization decision I made.
With no product manager on the project, I covered three roles at once: as design lead, every UI state across the full weekly lifecycle; as product manager, the PRD, the requirements and edge cases, and coordination across clinical, engineering, and leadership; as data scientist, the training data the AI needed to hold a goal-setting conversation at all.
When the clinical team's bandwidth for training sessions ran out, I didn't escalate — I generated the training data myself, grounded in real patient–clinician visit transcripts. We launched with physical activity only to constrain scope and QA, with a clear plan to expand. It wasn't in my job description. It was what the project needed.
The design covered the full weekly lifecycle — empty state, the goal-setting conversation, daily tracking, a mid-week check-in from Allie, and an end-of-week review that celebrates or gently regroups. On the clinician side, goals appear in the Foresight Care Plan alongside measurements and medications — clinically equal, not secondary.
Real usage kept teaching us. Some patients froze at an open-ended conversation, so I added suggestion chips drawn from their own history and medical context — still fully conversational, just less blank canvas. And to close the loop with in-person care, I designed a widget that reads visit transcripts and surfaces the goal a patient just discussed with their clinician — on their home screen the same day.
Active goal-trackers grew week over week from launch onward — patients weren't just setting goals, they were coming back to track them. Clinicians reported that patients were arriving at visits with goals already set. That shifted the nature of the appointment — instead of spending 5–7 minutes creating a goal, they could spend that time discussing how the goal was going.
That was the original ask. That's what shipped. The specific numbers aren't public — I'm glad to walk through them in person.
After leaving Codex, I rebuilt the SMART Goals experience as a fully functional prototype — live suggestion chips, free-text input, a working Allie holding a real conversation on the other end. Built with Claude Code. Not a click-through: it works.
Ask me for a walkthrough — I'd rather show than tell.
