Built a novel modular compressed images format, achieving a 7x decrease in dataset size, while maintaining accuracy (±1%) in image classification tasks. Engineered a column store for our image format exploiting the triangular nature of the data to achieve a fast, and shareable storage that reduced dataset size by 2.2x and reconstruction cost by 40%. Accepted as Frequency Store at CiDR 2025.
Head Teaching Fellow for CS205 - Graduate level course in High Performance Computing. Created homework problems, conducted office hours, and led sections.
Worked on Ahead of Time inductors for inference in Nequip, enabling constant folding for weights, leading to 40% faster inference.
•
Built a modular, distributed github CI test suite for Allegro, reducing testing time by 2x and automating the transition between development and production repositories
•
Modeled Gautz’s ACE descriptors using C++, KOKKOS, and CUDA handling 2nd order derivatives for FLARE, an open-source package that uses Active Learning to create fast interatomic potentials.
•
Contributed and co-authored to a paper presented at APS titled "High-performance training and inference for deep equivariant interatomic potentials". Submitted to Royal Society of Chemistry
