Relevant Experience & Education Highlights

Demonstrates deep expertise in developing compilers and kernels for AI hardware architectures, including GEMM, convolutions, and attention mechanisms, with strong alignment to d-Matrix's requirements for software kernels on next-generation AI compute engines.

1. Development, enhancement, and maintenance of software kernels for next-generation AI hardware

Built and optimized kernels for AI accelerators, GPUs, and specialized processors.

• Performed HW/SW co-design for AI architecture at a major semiconductor company, creating data-movement compiler maximizing memory-hierarchy efficiency for high-throughput GEMM, convolution, and multi-head attention workloads.
• Developed flexible attention compiler supporting arbitrary tensor shapes, masks, padding, reordering, layer fusion, and software pipelining for VLIW-SIMD cores.
• Optimized GPU kernels using CUDA and HIP for finite volume simulations, implementing operator fusion, loop flattening, tiling, achieving 10x performance on single GPU and 20x on multi-GPU setups.
• Implemented wrappers for sparse linear solvers across Linux, Mac, Windows using Clang, MPITrampoline, and MPI at a national research laboratory.

2. Strong understanding of hardware architectures and mapping algorithms to architecture

Mapped computational graphs and algorithms from AI frameworks to diverse hardware targets.

• Optimized deep learning models for neutronics simulations using Torch Compiler and TVM at a national laboratory, reducing memory usage and inference time, benchmarking across 200+ isotopes.
• Developed software pipelines for reactor simulations launching 4000+ instances, implementing MPI and NCCL-based parallel algorithms for generative AI data processing.
• Built ML library for time series classification with time series forest and KNN algorithms during Google Summer of Code, achieving 6x speedup via data randomization and unified APIs.

3. Full-stack toolchain, compiler infrastructure, and hardware-software co-design

Collaborated across teams to optimize compilers, validate silicon, and deploy production AI models.

• Collaborated with architecture, emulation, quantization teams at a major semiconductor company to validate silicon, implement inference flows, and deploy production-grade AI models.
• Developed full research proposal as principal investigator for large language model optimizing training with ZeRO-Infinity and DeepSpeed at a national laboratory.
• Created RangeEnclosures.jl Julia package and range over-approximation algorithms improving bounds by 70%, implementing branch and bound for polynomial bounding at NumFOCUS.
• Implemented Julia packages for range enclosures and branch-and-bound algorithms minimizing relative errors in mathematical modeling.

Requirements & Candidate Alignment