Things I've built, explored, and am still figuring out.
A personal portfolio and blog born from curiosity. A space to share ideas, explore the intersection of AI technology and human values, and think across disciplines like psychology, data science, and beyond. Built to reflect how I see the world: interdisciplinary, evolving, and rooted in what matters.
A dashboard for tracking job applications with filtering, sorting, and calendar views. Features search, status chips, and a monthly calendar for deadline tracking. Currently private while optimizing for efficiency.
Applying causal inference methods to evaluate policy interventions using observational data. Exploring treatment effect estimation with double machine learning and instrumental variables.
Building NLP pipelines for text classification and topic modeling. Working with real-world corpora to extract structure from unstructured text using transformer-based architectures.
Fine-tuning a BERT model to classify text messages sent during natural disasters into categories like medical help, water, food, and shelter. Designed to help emergency responders quickly triage incoming communications and prioritize relief efforts.
Spatial data analysis combining satellite imagery, geographic information systems, and statistical modeling to study urban patterns and economic geography.
Investigating what predicts who gets misclassified by the COMPAS recidivism tool, and whether those patterns differ across racial groups. Built on ProPublica's dataset (n=17,479) from Broward County, FL. Found that age x race interaction is the strongest predictor of misclassification, with error types splitting along racial lines. Extended with GAM analysis, separate FP/FN modeling, and XGBoost with SHAP interpretability.
Solving the Unit-Profit Minimum Knapsack Problem with compactness constraints using two approaches: a Mixed-Integer Programming formulation (Gurobi) for exact optima, and a custom O(n·t·Δ) dynamic programming algorithm for the decision version. Generated multi-modal test instances at multiple scales, validated DP correctness against all MIP solutions, and benchmarked performance.