Three years ago I was a hematology-oncology fellow in Detroit, and I could not write a single line of code. Today I build software that oncologists use to learn and practice. I did not take a sabbatical, go back to school, or quit medicine. I learned in the margins, the way most of us learn anything after residency: a little at a time, between clinic days, mostly by breaking things.
I want to share what that taught me, because more physicians are trying this now than ever. AI coding tools have made it possible for a doctor with no computer science background to build a working app in a weekend. That is exciting. It is also a little dangerous, and we should talk about both.
Why I started
I did not start coding because I loved technology. I started because I kept running into problems in my own training that nobody was going to fix for me. Oncology moves fast. New approvals, new trials, new guidelines. As a fellow, I felt the gap between what I needed to know and the tools I had to learn it. I assumed someone in industry would eventually build the right thing. Then I realized the people who understand the problem best are the ones living it.
That eventually became MeDucation AI, a tutor for oncologists. But the product matters less than the lesson: Physicians are not just users of technology. We can be builders.
Real physician voices, twice a week
Free, and one click to unsubscribe.
What I got wrong
I made almost every mistake you can make.
I built features nobody asked for. I spent weeks on things that looked impressive and solved nothing. The fix was a question I now ask about anything I build or evaluate: Does it save lives, save time, or save money? If it does none of those, it does not matter how clever it is.
I trusted the AI too much. When I started testing large language models on oncology questions, I found patterns that worried me. Models performed worse on less common presentations, such as male breast cancer. They also tended to latch onto a specific mutation named in a question and jump straight to the associated treatment, instead of reasoning through the whole case. A tired clinician would not catch that. A model that sounds confident is not the same as a model that is right.
I underestimated how slowly doctors adopt new tools. Physicians need much more time to trust a tool than the general public does, even when they can see it would help them. That is not stubbornness. It is the right instinct for people whose mistakes land on patients. If you build for doctors, you have to earn that trust, and you cannot rush it.
What doctors should build, and what they should leave alone
Here is where I have landed.
Build tools for your own workflow and your own learning. Calculators, teaching tools, templates, ways to organize information you already understand deeply. You are the best person in the world to judge whether these work, because you use them every day.
Be very careful with anything that touches patient data or makes clinical decisions. A weekend app does not have the privacy protections, validation, or oversight that patient care requires. Drug interaction checking is a good example. Every documented interaction is in the label, so automated tools flag everything. Most of those flags have no real clinical consequence. A few do. Pharmacists and prescribing physicians know the difference from experience. A tool you built over a weekend does not.
Know when to hand off. Building a prototype is one skill. Running safe, reliable software for other people is a different one. There is no shame in taking an idea as far as you can and then partnering with people who do this full time.
How to start
You do not need a degree. You need a problem you care about and a few hours a week.
Pick one small, annoying problem from your own work. Use the AI coding tools that exist now, but read what they produce and ask them to explain it, so you learn instead of just copying. Ship something tiny and show it to one colleague. Then fix what they hate about it.
Expect it to be slow. I still have only a couple of hours most days outside clinical work. That has been enough.
Why this matters
The next generation of clinical tools will be shaped by whoever shows up to build them. If physicians are not in that room, the tools will be built around what is easy to measure instead of what actually matters at the bedside.
You do not have to become an engineer. But understanding how these tools are built, and where they fail, may become as basic to practicing medicine as reading a trial. I learned that by building. I suspect many of you could too.
Roupen Odabashian is a medical oncologist in Canada.
