Medical education apps, built by a doctor
WardRun, Medacted and Resus One help medical students and junior doctors learn a few minutes a day. I’m a doctor, not an engineer: I built them with AI tools to see how hard it is to make the software, not just use it.
What I actually did
In 2026 I took medical education apps of my own from an idea to the App Store, using AI tooling to build them. I didn’t write the code. What I did was everything around it: decide what the software should do, write the clinical content, test it on real phones, get it through review, and, most of all, catch the errors the models made. There were a lot, and they were convincing.
I now know enough about how the software is made to see where it’ll go wrong, and enough medicine to catch most of it when it does.
The apps
Three tools for medical students and junior doctors. Everyone gets the same questions each day, so scores mean something. They’re free to use, and WardRun has a paid study mode for revision.
WardRun
You get one shift a day: three wards and one attempt, with questions mapped to the UK Medical Licensing Assessment. Leaderboards by medical school, so you can see where yours stands. There’s a paid study mode for revision.
Medacted
A disease file with every word blacked out. Guess words to uncover it, then name the disease. Every file is checked against national guidance. Good for revision because you end up reading the whole thing.
Resus One
A patient, some notes, a set of obs, and a registrar who wants an answer. Decide, then see what the guideline says. For foundation doctors and final years.
INSET, NORTH and Crown v King are puzzles about geography and criminal law. I built them because a daily puzzle is the simplest product there is, and I wanted to learn the route from idea to App Store before I trusted myself with anything clinical.
Questions about the apps go to playwardrun@gmail.com.
How it works, and what went wrong
I decide what the model writes, what it’s checked against, and what ships. Claude Code writes the code. I read the diffs. Clinical content gets a validator, a second set of agents checking it against NICE, UKHSA, CDC and WHO, a third pass trying to break the second, and me at the end. Two things still went wrong, and they’re the useful part.
The App Store rejection
The reviewer reported that the app froze on the paywall. It hadn’t. On an iPad with no haptic motor, the only feedback for a failed purchase was a vibration, so the reviewer tapped, felt nothing, and had no way back. It took a long time to reproduce because I was testing on a phone. The fix was small.
The videos I had to pull
I built a pipeline that rendered a batch of short revision videos, each opening with a run of quick cuts. Every one of them broke the flash threshold in WCAG 2.3.1, the limit that exists for people with photosensitive epilepsy. Some had already been posted. I caught it by watching one back, because every check I had tested the pipeline rather than the finished file. Now a script decodes each rendered video and fails it if the flashes are too close together. Both times the fix was the same: test the finished thing, on the device it will actually run on.
Building something with AI for patients or clinicians?
Or making something about AI in medicine. I’m happy to talk, as an adviser or a board member.