T2D Risk Screener
A calibrated diabetes pre-screening tool trained on Korean NHIS health-exam data (about 1 million patients). Flags patients who likely need a glucose test, with a SHAP explanation behind every score.
I build ML systems end to end: data pipeline, model, evaluation, and deployment. Recent projects span gradient-boosted risk models, transformer NLP, and computer vision. BS in Computational and Data Sciences (summa cum laude), based in Seoul.
What pulled me into data science was watching AlphaFold's CASP14 result: a model solving a structural biology problem people had worked on for decades. In my own work, I care most about the parts that decide whether a model is actually useful in production: honest evaluation on held-out data, calibrated probabilities, explainability, and a real deployment instead of a notebook that only I can run. I'm a Computational and Data Sciences graduate of George Mason University Korea (summa cum laude, GPA 4.0/4.0), based in Seoul. I'm a native English speaker with business-proficient Korean and a former U.S. Marine Korean linguist, and I've worked professionally in Korea for several years. I'm open to roles across data science, machine learning engineering, data and backend engineering, and technical product management, in any industry, with a particular interest in healthcare and biomedical work.
A calibrated diabetes pre-screening tool trained on Korean NHIS health-exam data (about 1 million patients). Flags patients who likely need a glucose test, with a SHAP explanation behind every score.
An ablation testing whether biomedical pretraining actually helps on adverse drug event (ADE) screening. It does, modestly: PubMedBERT reaches 0.938 macro-F1 vs. 0.920 for base BERT, with paired-bootstrap confidence intervals separating training noise from test-set noise.
A fine-tuned image classifier that predicts a photo's country from street-level imagery. On a shared 1,000-image benchmark it beat a human comparison group by about 6× (64.7% vs. 10.85% average, McNemar χ²=444.7, p<0.001).
A Python package that takes a messy edge table to an interactive Plotly chart: clean it with a report of every change, build a NetworkX graph, and plot it with one set of arguments. Version 1.0 is a ground-up rewrite, tested at 99% coverage, with FAOSTAT trade data as the worked example.