I build products clinicians trust, and decide where AI helps and where people stay in the loop.

01

Recommendations with receipts: AI learning plans clinicians and their institutions can trust

Taking LLM features from prototype to a paid product for physicians, nurse practitioners and PAs, with clinical review of model output and clear points where people stay in the loop.

The problem

A library of 500+ accredited courses only helps if a busy clinician can find the right piece at the right time. Generic "you might also like" suggestions don't earn trust in a clinical setting. Clinicians want to know why something is in front of them, and the institutions training them want to act on it.

The foundation came first

Before LLMs could organize content, I built a tagging framework mapped to board certification blueprints and an internal medical catalog that organizes the full library for search, reuse and machine learning. That structure is what the AI features stand on today.

The calls I made

1. Show the reasoning, not just the result.Each recommendation comes with a generated explanation of why it was chosen. That's partly an interaction design problem (where and how the reasoning appears) and partly a system behavior problem (the model writing each explanation).
2. Review the output like a clinician would.Reviewing model output myself, I found recommendations had to draw on the correct answer and its explanation, not just the question stem. I set up a human recommendation set to compare against, and early examples showed why it matters: the model matched content to the likely diagnosis, while a clinician matched it to the symptom workup.
3. Rate usefulness item by item.I co-wrote the testing plan with our clinical lead and pushed for a usefulness rating on every recommendation, not one score for the whole set. A random suggestion can hide inside a good average; item level ratings expose it.
4. Evaluate before anything ships, and write down where humans stay in the loop.Test sets and scoring rubrics run every time the model or prompt changes. Guardrails spell out which outputs need human review, so that line is a decision, not an accident.
5. QA at the level a clinician would notice.Testing the engine that links recommendations to content, I caught mapping errors that automated checks missed: the right lecture, but the wrong segment of it.

Where it landed

AI learning plans, generated from each learner's assessment results, went live in October 2026 as a paid feature for institutional customers, our first user facing AI product. Within weeks a customer asked for exactly this feature on their own. Early clinical feedback says the recommendations are improving and the explanation language still needs work, which is what we're iterating on now.

From recommendation to action

A plan only matters if an educator can assign it. I designed the end to end flow for turning AI learning plans into assignments myself, prototyping with AI design tools, and grounded it in production data rather than assumptions:

02

Retiring a ten year old platform: nine products moved, no data loss, no downtime

I led the strategy, sequencing and cross functional execution that moved every product off a legacy platform onto the new learning platform I'd helped design, carrying clinicians' CME history, content, bookmarks and comments along with them.

Why it was hard

CME credit is what clinicians need for licensure and recertification, so a migration that loses or miscounts it isn't a bug, it's a broken promise. Meanwhile the business kept launching products on both platforms, and the team changed several times over the years it took.

"I need to balance [two product lines] and CME carefully to make sure there are as few waves as possible."My planning note, 2019

The calls I made

1. One product with filtered views, not forked products.When two exam audiences needed different slices of the same content, I built the new platform with a filter system instead of splitting into separate products. One thing to maintain, and each learner sees only what's relevant to their test.
2. Sequence for the fewest waves.Migrations were ordered to protect CME accuracy and limit how many times any customer felt a change.
3. Call the slip early and restructure.In early 2024 every migration was off track. I said so to the CEO that day, reassigned ownership, took technical delivery on myself, and hired a technical project manager. We built a feature parity list, ran a weekly migration sync, and finished.
4. Don't migrate what isn't worth migrating.Low value legacy data was exported and archived instead of moved. The team called it the "pizza party" sunset.
5. Close the loop after cutover.After the last move I checked that engagement data for migrated products still flowed into our dashboards, so the analytics stayed continuous.

Where it landed

The first product went live on the new platform in March 2021, alongside a new native listening app. The last legacy product moved in June 2025 and the old platform was retired. Three months later every course was available in the native mobile app. Along the way we retired older products with minimal customer friction.

Through an acquisition

Earlier, I led operations through the company's acquisition by a larger public education company.

03

Designing for skeptical buyers: an Urgent Care line built on research

I launched an Urgent Care education line for individual clinicians and the centers that employ them, then used research to design an institutional offering that directors would actually adopt.

How I approached it

I interviewed urgent care center directors and providers, wrote a press release before building anything, mapped journeys for the provider and the medical director, and set explicit goals and non goals.

What the research showed

Directors wanted plug and play education and analytics, but were skeptical and hesitant to adopt formal assessment exams, let alone enforce them. Providers would take part as long as the process wasn't punitive.

The calls I made

1. Design for the buyer and the learner separately.Directors needed reporting and accreditation compliance; providers needed something that didn't feel like a test they could fail. The product had to give each what they needed.
2. Low stakes instead of formal assessment.The design became short monthly modules of five questions, rolling up into an admin dashboard and compliance reports for the center.
3. Partner with the field.We co-branded the program with a national urgent care professional association.

Also

A mobile app turnaround: 3.7 to 4.6 stars on iOS, 2.8 to 4.2 on Android

I made the mobile app the top product priority, turned around poorly rated iOS and Android apps, and finished a unified app vision I'd written years earlier. Ratings below run from October 2022 to November 2026.

1. Write the vision early, keep the car pointed at it.In 2019 I wrote a spec for one app covering audio, video, question banks, CME and accounts, deliberately leaving room for later features. The first version launched in March 2021; profiles, group assignments and a daily plan all shipped later on that foundation.
2. Fix volume before score.Low ratings were partly a sample problem. We set a goal of 500 ratings averaging 4.5 or better, showed an in app rating prompt to 5% of users per platform behind a feature flag, and fed what they told us into the backlog.
3. Staff it.I hired a dedicated native developer, then shipped native question banks, offline video, playback and performance fixes, and finally every course in the app.