Inspiration --> House MD and doctor diagnosing
incident.io's brief says most AI answers after the fact. Medicine is the opposite: a patient doesn't wait for your prompt. One of us studies biomedical engineering, and clinical reasoning is still mostly taught from static text, where you read the answer and never see what happens when you get it wrong. We wanted to learn it the way a House episode plays out: a new symptom every scene, the wrong treatment making things worse, and a colleague arguing about the diagnosis with you.
What it does --> anatomy treatment simulator
Anatomy Park MD is a live medical mystery. You're the lead doctor and your patient has just collapsed.
- Examine what a human would see, like a drooping face, a bruise or a rash, then scan six layers of the body underneath.
- Page your fellow, Dr. Foreman. Time stops while you talk. He only knows what you tell him, builds a differential and offers treatments.
- Prepare a treatment and physically deliver it: drag the syringe in, charge the paddles and shock, cut with the scalpel. Or type anything you want to try, and it appears in your hands.
- The body reacts: cured, improving, no change, worse, much worse or lethal. Amputate a limb and it stays gone.
- Every case ends with an incident post-mortem: root cause, timeline and what you made worse.
There's no multiple choice and no answer key. Educational game, not medical advice.
How we built it
The browser (JavaScript, Vite, SVG and Canvas) runs everything you see and touch. A small FastAPI server talks to Claude through five endpoints, and every call uses tool use with a fixed schema, so a reply can't break the game.
- Claude Opus writes each hidden case, and the server keeps a pool ready so starting a game is instant.
- Foreman runs on Claude Haiku and never sees the diagnosis, so he can be confidently wrong.
- A separate Claude Opus physiology judge knows the truth and decides what each treatment really does to this patient.
- Our own simulation turns problems at 17 body sites into heart rate, blood pressure, oxygen and the ECG, every frame. Claude decides what goes wrong and our code moves the numbers, so the monitor can't hallucinate a heart rate.
- Claude maps each new symptom onto 22 visual effects that our renderer draws and scales with live severity, and it draws an item for any treatment you invent.
Challenges we ran into
- We pivoted on the evening of day one, after deciding our first idea wasn't AI-driven enough, and rebuilt from scratch.
- A single AI that knows the diagnosis leaks it, so we split the knowledge across two agents.
- Latency: writing a case takes about 25 seconds, which kills a live game. We pre-generate cases and freeze time while the AI is thinking.
- Generated images were too slow and inconsistent for live symptoms, so Claude picks from a fixed vocabulary of effects and our own renderer draws them.
Accomplishments that we're proud of
- A fellow who can be honestly wrong in a world that's honestly right.
- A patient with permanent consequences. In one run we amputated a healthy arm, the patient went into septic shock, and Foreman told us off and changed his diagnosis.
- Physical treatments instead of menus: every delivery type has its own action.
- A post-mortem that reads like a real incident report.
What we learned
- What an agent is not allowed to know can matter as much as what it knows.
- Fixed schemas make model output safe to build a live game on.
- Letting AI decide what happens while deterministic code handles the moment-to-moment numbers gives you both creativity and reliability.
- Patients and production systems fail in surprisingly similar ways: alerts, triage, fixes, cascades and post-mortems.
What's next for Anatomy Park MD
- Voice, so you can talk to Foreman out loud.
- Foreman interrupting you on his own when the monitor alarms.
- Teaching cases built with and for medical students.
