Developing a GNN-based framework informed by curvature of directed graphs, integrating optimal transport theory for graph learning tasks. Research at Harvard SEAS with Prof. Melanie Weber, exploring geometric learning on directed networks.

Built a machine learning model to infer outdoor time using light exposure data from Axivity AX3 sensors; enabled large-scale behavior analysis. First-author research conducted at Oxford University Big Data Institute with Prof. Aiden Doherty. (To be submitted 2025).

Designed a novel optimization framework that boosts CNN-based dense object detection using annealing-inspired heuristics for better convergence and accuracy. Co-first author research for Harvard Advanced Computer Vision, in collaboration with Trace Baxley. CS2831.

We show that using Graph Neural Networks to both warm-start the Ford-Fulkerson algorithm and guide augmenting path selection significantly reduces the number of iterations required to reach max-flow. Our hybrid method leverages GNN predictions for flow initialization and edge prioritization, with theoretical guarantees relating prediction accuracy to algorithmic speedup. With Trace Baxley. CS2241.

We train neural networks to approximate the swept volumes of moving shapes using signed distance functions (SDFs), capturing how objects evolve over time along specified trajectories. By experimenting with network architectures, we address challenges posed by discontinuities and complex motions in predicting accurate SDF representations. With Kimberly Herrera and Juan Parra.