Summary
Software Engineer and Computational Scientist with a PhD in Applied Mathematics and 7+ years of experience designing high-performance computing systems, low-latency runtime components, and massive-scale simulation engines. Experienced in JAX/XLA runtime and ML infrastructure (data pipelines, embedding serving, distributed training at cross-datacenter scale), low-latency C++ systems and GPU acceleration (CUDA, XLA), and large-scale numerical solvers. Proven track record of optimizing core compute infrastructure to drive significant cost savings and throughput increases in production environments.
Work Experience
Google LLC
Apr. 2023 – Present
Software Engineer – Machine Learning Infrastructure | Sunnyvale, CA
-
Data Infrastructure (PyGrain & JAX Data)
- Maintain and scale PyGrain and JAX Data, the foundational data pipeline libraries handling ingestion and processing for JAX-based training workloads globally across Google. Designed and built the PyGrain autotuner, a closed-loop online controller using Universal Scalability Law (USL) and Markovian queueing models to dynamically tune worker concurrency, prefetch depth, and batch parameters, significantly reducing the time to build high-performance data pipelines.
-
JAX CPU Embedding Library
- Designed and built Google's high-performance native C++ embedding library from the ground up, featuring cache-optimized embedding table layouts, lock-free lookup structures, and lookup coalescing, delivering a 2x to 4x throughput improvement over the incumbent TensorFlow-based solution for embedding serving.
-
JAX Parameter Server Training (PST)
- Engineered CPU-based distributed training architectures optimized for massive-scale recommendation and ranking models. Optimized parameter server infrastructure to deliver up to a 3x throughput improvement in dense worker/server topographies by implementing copy-eliding RPC handlers, custom cache-friendly C++ data structures, and low-overhead threading models. Developed a scalable native C++ optimizer library from scratch with horizontally scalable routines tailored for these topologies, and partnered with cross-functional stakeholders (Ads, YouTube, Play, Geo) to successfully deploy PST into production.
-
Megascale XLA
- Contributed to the Megascale XLA team, implementing topology-aware collective communication primitives within the XLA runtime, including hierarchical tree-based collective planning across intra- and inter-datacenter network tiers to enable efficient distributed training at cross-metro scale. Received the internal “Gold Perfy” Award for this contribution.
Aurora Innovation
Oct. 2021 – Apr. 2023
Senior Software Engineer – Motion Planning Simulation | Mountain View, CA
- Architected core high-performance software components in C++ for a highly distributed, massive-scale motion planning simulation engine executing hundreds of thousands of autonomous vehicle scenarios daily.
- Developed dynamic agent behavior routing frameworks, introducing adversarial interaction profiles and real-time lane-changing mechanics to evaluate safety-critical edge cases.
- Created automated validation tools for Foreign Object Debris (FOD) simulations, engineering workflows that dynamically combined real-world sensor logs with synthetic environmental augmentations.
- Discovered and executed sweeping runtime optimizations within the mapping and routing engine, generating audited infrastructure savings of ~$130,000 per month in cloud compute expenditures.
Siemens PLM Software
Jul. 2019 – Oct. 2021
Advanced Software Engineer – Physics Solvers | Lebanon, NH
- Maintained and optimized parallel fluid dynamics solvers within the flagship industrial simulation suite STAR-CCM+.
- Researched and implemented highly parallelized computational methods (MPI, OpenMP) to resolve large-scale, complex partial differential equations on multi-node cluster topographies.
- Researched and developed GPU offloading strategies for core Algebraic Multi-Grid (AMG) linear solvers using CUDA.