Research

During my undergraduate studies, I worked on reducing the overhead of differentially private ML training by investigating GPU microarchitectural and kernel-level optimizations for gradient clipping in DP-SGD — an operation that contributes to the algorithm's 1.5×–40× slowdown over standard SGD.

Through this project, I witnessed firsthand how emerging data intensive applications — particularly those with privacy requirements — expose significant hardware-software mismatches that result in avoidable overheads. As Moore's Law fades, there is a growing need to examine how the systems stack aligns to emerging workloads, from microarchitecture to OS and networks. My research interests lie in optimizing and co-designing these systems layers to better suit modern computing, from both performance and security perspectives.

Currently, I am working on optimizing memory exceptions.