Samir C. Grover, M.D.

We built this delightful intro to #endoscopy simulator for the entry endoscopy trainee in three.js with 45 minutes of vibe coding.

Crazy what you can build rapidly (and disposably) in all the frontier models -> rendering software.

It is quick and dirty and flawed, but most importantly can be made and deployed in almost an instant.

have a peek

groverlab.ca/endoscopy_simu…

#MedTwitter #gitwitter

I built a dashboard for my AI agent stack then rapidly got overwhelmed.

Instead of tracking agents, what I do now is
1/ Observe AI agents
2/ Entrust them gradually - akin to how we entrust residents in medical training - when they prove they know what we can do
3/ Then have them present to me AFTER the fact - and they tell me what they did.

Allows for more autonomous work, and saves me the discipline of looking at a dense dashboard.

I describe my multi-harness stack and how it works in this Substack article (along with some nerdy med-ed stuff about how this is the same as competency-based education...for AI agents)

samirgrovermd.substack.com/p/i-stopped-wa…

There is an epidemic of non-introduction in healthcare!

Went with dad to hospital for the past 3 days. People expecting to poke and prod him and give ask him personal questions and give him info... With no indication of saying who they are, what they do!

A smaller subset even got flustered or even upset when asked who they were and why they were there.

Across all clinical care groups.

I didn't like it at all. It was almost dehumanizing.

Yes! We should normalize walking pads on virtual meetings and appointments!

Within three years, I hope EVERY simulation centre has a small server.

It wont be connected to the cloud. It will run vision and language models locally, review clinical and procedural video as its recorded, generate practice cases and exam items, and never send learner or patient-adjacent data out of the building.

I think this is where health professions education is heading, for three reasons: (1) privacy; (2) reproducibility of open-weights models and (3) cost (the 100th learner and the 1st learner cost the same)

Small models aren't ready for everything. They still make things up, and their safety has barely been studied. But I posit they still play a valuable role for health professions education.

Full piece on Substack:
In praise of small: why health professions education needs local AI
samirgrovermd.substack.com/p/in-praise-of…