Machine Learning Researcher, Michigan State University
I'm a machine learning researcher in the Medical Imaging and Data Integration Lab at MSU's Institute for Quantitative Health Science and Engineering. Our group works with CDC/NIOSH on screening chest X-rays for pneumoconiosis, the lung disease miners and industrial workers get from breathing coal and silica dust.
My part is adapting vision and multimodal foundation models to this problem, and building the offline desktop application that lets radiologists and B readers actually use them. I also mentor undergraduates in the lab.
I finished my B.S. in Computational Data Science, with a minor in mathematics, at MSU in 2026. Alongside my degree I spent three years at iCER, MSU's HPC center, writing tools for people who run jobs on the cluster. On the side I'm co-founding DockPilot, which makes scheduling software for warehouse loading docks.
Deep learning models for screening and grading chest X-rays performed comparably to certified B readers on in-distribution test images. Out-of-distribution cases remain harder.
One of the first studies to apply deep learning to multiclass grading of occupational lung disease, treating the four ILO grades as an ordered scale.
Compared six loss functions for ordinal grading on 2,077 NIOSH radiographs. CORN loss did best, at 70.9% accuracy and 89.8% specificity.
The job-script checker described below, presented with a live demo.
A desktop application that runs our models entirely offline, so patient images never leave the reading workstation. I lead development: the PySide6 interface, ONNX inference, about 1,860 automated tests, Windows CI, and the installer. It's in beta with radiologists now.
Fine-tuning and benchmarking vision and multimodal models on our NIOSH data. Held-out accuracy went from 75% to 85% (ROC-AUC 0.80 to 0.96), and LoRA on MedGemma 4B went from 82% to 90%. I'm now working on uncertainty-aware labels for films where expert readers disagree.
Catches mistakes in SLURM scripts before they're submitted, using rule-based checks, ShellCheck, and a local LLM for the fuzzier cases. Deployed on Open OnDemand for MSU researchers. I also built a model that suggests CPU, memory, GPU, and wall-time requests from past job data.
A local retrieval system over PDFs and Word files. It answers questions, quotes the passage each claim comes from, and declines when it can't find support. Keyword search in PostgreSQL plus vector search in Qdrant, with a local model through Ollama.
Outside the lab I co-founded the Polish Student Association at MSU, which reached 250 members in its first year, and was a Resident Assistant for a hall of 100+ students. I'm a native Polish speaker.