Accelerating multi-GPU distributed training with NVSHMEM-optimized fine-grained communication.

Currently making 4D Gaussian Splatting faster by leveraging VLMs, with the Interactive 3D Vision & Learning Lab. Previously, developed an adaptive AR labeling tool in Unity C# and built software enabling navigation through 20+ realistic testing environments for a user study [paper].

• Computer Networks (Spring 2026) – Helping students write IP and TCP from scratch.
• Graduate Deep Learning (Fall 2025) – Helped students figure out transformers, autoencoders, diffusion, and more!
• Computer Graphics (Fall 2024) – Helped students trace rays and fight OpenGL, for the longest running computer graphics course in the known universe.
• Computer Vision (Spring 2024) – Helped students understand concepts from traditional computer vision to modern deep architectures.

A static analyzer for PyTorch that infers tensor shapes and catches dimension mismatch errors before runtime. Leverages shape annotations around function params and polynomial canonicalization to check constraints without an SMT solver. Still in development!

A fast, action-conditioned next-frame generator for Minecraft, inspired by Oasis and the MAR framework for image generation. Designed for model-based control, it enables ~4× faster sampling than full-frame diffusion methods.

An educational cryptocurrency inspired by Bitcoin, featuring proof-of-work mining, transaction verification, and fork resolution. It simplifies blockchain mechanics using global balance tracking instead of UTXOs, and runs over a custom virtual OS and peer-to-peer network for decentralized coordination.

A from-scratch stable diffusion model for conditionally generating street view images from a corresponding satellite image. Given a satellite image, the model is able to generate a plausible looking view from the ground.

A simulation modeling the transport networks generated by Physarum (a slime mold). Implemented with Unity and powered by a compute shader, capable of individually updating the dynamics of 1 million+ interacting slime agents in real time.

A better Word Hunt! Developed with React and playable here! Click tiles to select them and type to place a letter. After creating your desired shape, press randomize to switch things up and drag together adjacent letters to form words. Longer words = more points!

A set of models for predicting country and coordinates from a given street view image. Scraped 50,000+ images from Europe using Selenium + Google StreetView API, and separately fine-tuned ConvNeXt and BiT for country and real-valued coordinate predicitions. Achieved 61% country classification accuracy (seems low, but you try it with a single street view image!) and 60% accuracy on coordinate prediction within 500km.