Projects

AI-powered course discovery and graph-based learning exploration using MIT OpenCourseWare. Semantic search with prerequisite expansion and personalized learning path generation.

Pipeline for synthetic aerial stereo dataset generation, stereo depth estimation, and point cloud semantic segmentation. Uses procedural BlenderProc rendering and GTA V capture to produce stereo RGB, depth, disparity, and segmentation ground truth. Fine-tunes RAFT-Stereo for disparity and trains a sparse voxel UNet on the back-projected point clouds.

Custom 3D deep learning architectures for LiDAR point cloud segmentation and individual tree detection in forestry. Trained on cloud GPU systems. Part of a joint research initiative with MIT and the University of Idaho on Geiger-mode LiDAR applications in forest mensuration.

Improves the best known upper bound on the asymptotic square packing problem — a 50-year-old open problem of Erdős on how efficiently unit squares can be packed into a large square.