Mariusz Kurman
Physician, ML engineer, language-model tinkerer.
Hi, I’m Mariusz. I build and train language models, write tools for agents, and work on clinical AI. I studied medicine and engineering, and somehow ended up spending a lot of time thinking about convolutions inside language models.
I’m interested in what happens below the API: architectures, training data, optimization, and post-training. Much of my model work is self-directed. I like being able to change the model itself, run the experiment, and see where the idea falls apart.
Medicine is the other thread. I want agents that can gather the right evidence and reason about the whole patient—not just name the diagnosis. The goal is to give clinicians better help, not replace them.
Models & experiments
-
NeuroBLAST v3 — a 597M-parameter hybrid of attention and dilated causal convolution. I implemented it in JAX/Flax and PyTorch and pretrained it on about 144B tokens using TPUv4 through Google TRC. Checkpoint ↗
-
ConvGPT — what if the decoder worked in a much narrower space than the embeddings? Convolutional compression from 1,296 to 144 dimensions, about 250B training tokens, and custom vLLM/SGLang modules. Checkpoint ↗
The later ConvGPT-v2 is a different architecture: causal 1D/2D convolution, sparse retrieval, and custom Triton kernels.
-
Teacher-feedback GRPO — a training loop where a teacher scores student responses, with group-normalized rewards, a reference-policy KL term, and optional SFT on feedback. Released model ↗
These are experimental models, not claims of a new state of the art. Approximate training-token counts are documented in the linked model cards.
Clinical reasoning, beyond Q&A
I built ClinicalGym and ClinicalBench to let agents work through simulated patient encounters: ask questions, examine, order tests, diagnose, and record treatment decisions. The interesting part is the path they take, including what they miss.
They build on 1,300+ physician-authored cases, each validated by at least two other doctors. I wrote about 300 of them. The tooling records trajectories and exports preference and reward-model training data.
Other things I build
A bit of background
I lead AI/ML R&D at Kruk. Before and alongside that, I founded MedIT Solutions, built a team across engineering and medicine, and shipped Medcases. Earlier, I worked on ranking and author disambiguation at Research.com, and before that wrote Java and PHP.
I have an MD from the University of Warmia and Mazury and an engineering degree from Warsaw University of Technology. The full timeline is in my CV.
Say hi
If you’re working on language-model training, clinical agents, or something related, I’d be happy to talk. You can reach me at mariusz.kurman@gmail.com.