Research Intern

The year before, my grandfather had a stroke. He survived, but it took his speech and most of his recognition of the family. That's what put me on this project.

Research on stroke-risk prediction under Dr. Shalaginov, building neural-network models over clinical biomarker datasets.

I owned the data preprocessing pipeline — the unglamorous half, where most of the signal is won or lost — and built the clinician-facing web application that put the model in front of someone who could actually use it.

Three straight weeks of that summer went to a pipeline that kept dying on malformed records — the kind of debugging where every fix reveals two more. I came out with a lasting respect for anyone whose whole job is keeping data clean.

The hardest part wasn't the model. It was a single number: the decision threshold. Set it low and you flood clinicians with false alarms; set it high and you miss strokes. Someone has to pick, and the picking is a moral choice wearing a parameter's clothes.

The model reached 89.3% test accuracy. Digging into why, I found the training data underrepresented whole groups of patients — my first real lesson in asking "accurate for whom?". The work became a research paper, which I co-authored.