Helping make zkSync the world's first (as far as I know) GPU-accelerated zero-knowledge rollup.Helping Ethereum scale to thousands of transactions per second without sacrificing decentralization or security.Details:The zero-knowledge prover is compute intensive. Its cryptographic primitives (big-integer and elliptic curve point arithmetic, multi-scalar multiplication, number-theoretic transforms, hashing) benefit greatly from GPU acceleration.

Member of Pytorch team.I focused on making GPU training fast, numerically stable, and easy(er) for internal teams, external customers, and Pytorch community users.Among other duties, I:- Contributed Python frontends, CUDA kernels, and C++ backend fixes to improve mixed-precision, multi-GPU, and eager mode user experience.- Debugged, profiled, and optimized Pytorch networks.- Integrated new CUDA features with Pytorch.- Guided CUDA feature development to ensure upcoming features met Pytorch’s needs.

Ported performance-critical physics routines for rocket CFD code from Fortran to C++ using Kokkos, a library that enables the same code to compile for Nvidia GPUs, Intel Xeon Phis, or multicore vector CPUs. Optimized my code for all the above architectures, achieving up to 32X speedup over serial Fortran.Performed "code surgery" to interface Kokkos-accelerated C++ routines with large Fortran codebase, emphasizing minimal alteration to existing code and minimal impact on other developers. Achieved 3X speedup for overall code at multinode scale.Helped configure Xeon Phis on our cluster to allow multinode MPI jobs.

Adapted numerical techniques from barrier-crossing theory to polymer self-consistent field theory and other field-theoretic models. Devised parallelization and memory-saving schemes to optimize these techniques for various high-performance computing architectures using CUDA and MPI.