Inspired by Curriculum Learning and Probabilistic Roadmaps, PHO traverses a multidimensional homotopy space and discovers highly dynamic trajectories.
Designed an SO(3) equivariant autoencoder using spherical harmonics and a latent space traversal that separates rotation from deformation.
Reimplemented Decision Transformer replacing transformer with S4 model and demonstrated improved performance in credit assignment tasks.
Created a face search engine that uses a custom clustering method on ResNet vector embeddings (unsupervised). The UI is a scalable website.
Formulated time complexity verification of a program as synthesizing a fix-point function. The demo uses a custom language with Python syntax.
Educational tool to track student performance on projects, providing live feedback and tips, and automating infrastructure setup for instructors. Built with Python, DSPy, Kubernetes, and React; won Warp and Orbstack challenge prizes.
Developed AI agents for a multi-agent game, achieving 4th place in the competition. Implemented Huffman coding for cost-effective communication and utilized the A* algorithm for shortest pathfinding on a partially explored map.