Projects

A growing collection of my cool projects.

Work
project thumbnail

From Baseline to DeepSeek - Single-GPU MoE Training Efficiency

A systems-level analysis of training Mixture-of-Experts (MoE) Transformer models under single-GPU constraints. We compare naive PyTorch MoE, ScatterMoE, MegaBlocks, and DeepSeek-inspired architectures, revealing critical trade-offs between convergence behavior, memory footprint, and training throughput.
project thumbnail

SmallGraphGCN - Accelerating GNN Training on Batched Small Graphs

Discover how fused-edge-centric CUDA kernels dramatically accelerate Graph Neural Network training on molecular datasets. By rethinking parallelism strategies for batched small graphs, our approach achieves up to 3.1× faster forward execution and 1.3× end-to-end training speedup over PyTorch Geometric.
project thumbnail

Gaze-Guided Reinforcement Learning for Visual Search

Discover how gaze prediction from human eye-tracking enhances AI agents in object search tasks. By integrating visual attention into reinforcement learning through three novel methods, our approach enables faster, more effective navigation in simulated environments.
project thumbnail

Understanding Swap Regret 2.0

This blog post unpacks the "Swap Regret 2.0" paper through slide-by-slide insights, showing how the TreeSwap algorithm closes the gap between external and swap regret, advancing equilibrium computation in game theory and reinforcement learning.
project thumbnail

MTA Transit Time Prediction

Leveraging real-time data and machine learning to predict bus arrival times in New York City with route-based and grid-based approaches.

Fun