AI inference + high performance computing

Building high-performance systems for modern AI.

I am a Principal Software Engineer within Microsoft AI, specializing in inference optimization, GPU kernels, quantization, and model serving. I developed BenchPress, an agentic autoresearch system for autonomous inference-engine tuning across vLLM, SGLang, and TensorRT-LLM. Its agent diagnoses bottlenecks, designs experiments, and optimizes serving configurations against latency and throughput goals using profiling and reproducible benchmarks.

Previously at Microsoft AI, I optimized Triton kernels, integrated FP8 KV cache support, and built checkpointing and fault-tolerance infrastructure for large-scale training. Earlier, on Microsoft's DeepSpeed team, I optimized OpenAI DALL.E inference and added support for OPT, Falcon, and Phi-2. My earlier work spans edge AI at Blaize, Larq Compute Engine at Plumerai, and instruction-set and debugger engineering at Intel.

Experience

Microsoft AI

Principal Software Engineer

Developed BenchPress, an agentic autoresearch system for autonomous inference-engine tuning across vLLM, SGLang, and TensorRT-LLM. Built closed-loop optimization of quantization, speculative decoding, batching, scheduling, and KV-cache settings, guided by PyTorch and NVIDIA Nsight profiling and reproducible benchmarks.

Microsoft AI

Member of Technical Staff

Focused on inference performance for the Pre-training / Inference team. Optimized Triton kernels, integrated FP8 KV cache support, and designed checkpointing and fault-tolerance infrastructure for large-scale training.

Microsoft

Senior Research SDE · DeepSpeed

Optimized OpenAI DALL.E 2.0, 2.5, and 3.0 inference, significantly reducing latency and contributing to millions of dollars in service-cost savings. Extended DeepSpeed Inference with support for Meta OPT, Falcon, and Microsoft Phi-2.

Blaize

Staff / Senior Software Engineer (HPC/DL)

Quantized and optimized computer vision models including ResNet, OpenPose, U-Net, YOLO, and Mask R-CNN for the Blaize Graph Streaming Processor architecture.

Plumerai

Deep Learning Software Engineer

Led the design and development of Larq Compute Engine for mobile and embedded inference, using cache optimization, ARM NEON SIMD, multithreading, and hand-optimized assembly. Co-authored its MLSys 2021 paper and supervised three developers.

Intel

Software Engineer

Benchmarked Intel Deep Learning Boost AVX-512 VNNI instructions on Cascade Lake and implemented Control-Flow Enforcement Technology support in the GNU Debugger.

Research

High-performance computing, parallel algorithms, and computational fluid dynamics.

I received my PhD in Informatics and its Applications from the Technical University of Munich, with research in high-performance computing, parallel algorithms, and computational fluid dynamics. At the Institute for Advanced Study, I worked under the supervision of George Biros and Hans-Joachim Bungartz.

My doctoral work included TbSLAS, a distributed/shared-memory solver demonstrated on problems with one billion unknowns across 16,384 CPU cores.

A full list of peer-reviewed and pre-print manuscripts is available on my Google Scholar profile.