I build machine-learning and simulation systems for decisions that move people and goods. My work connects rare-event prediction, sequential decision policies, agent-based modelling, GIS, and applied statistics with careful validation and explicit operational tradeoffs.
I am a Transportation Engineering PhD candidate at Penn State, open to ML/AI engineering, research science, simulation, scientific programming, and applied-statistics roles.
- Hierarchical rare-event ML — improved rare-outcome average precision by 27.6% over a direct ExtraTrees baseline by modeling the response structure, with repeated internal validation and a semantic negative control.
- Equity-aware contextual bandits — a parcel-locker recommendation simulator that predicts out-of-home delivery propensity and chooses feasible access points from daily activity chains while balancing service value, pickup burden, and equity constraints.
- Parcel-locker freight ABM — a six-context NYC/Seattle model connecting synthetic e-commerce demand, OSM routing, MOVES4-derived emissions factors, a separately implemented private AERMOD stage, demographic exposure equity, and exact inference over a 2,700-run factorial experiment.
- Parcel-locker AnyLogic GIS — a last-mile digital twin that exposes the tradeoff between fewer failed deliveries and more personal pickup travel under higher locker adoption.
- Transit and health equity — a multi-city GIS and statistical pipeline connecting GTFS-based healthcare access, neighborhood inequality, poverty-group contrasts, and CDC PLACES outcomes.
- Bayesian spatiotemporal truck-crash modeling — sparse spatial adjacency, ST-ACF/STARMA diagnostics, and negative-binomial CAR–AR models in brms and Stan.
- City-scale OSM evacuation ABM — a collaborative project with Alex An, PhD, combining congestion-aware routing, spreading hazards, pedestrian behavior states, and scenario aggregation over OpenStreetMap-derived networks.
- Traffic-sign detection with YOLOv4-tiny — a brittle notebook experiment rebuilt as deterministic preprocessing, validated configuration, tests, provenance, and an honest model card.
Python · R · SQL · Java · scikit-learn · GeoPandas · OSMnx · NetworkX · PySAL · brms / Stan · AnyLogic · MATSim