Start with a representative workload
Freeze the input, output, correctness criteria, and timing boundary before tuning. A useful baseline must be reproducible.
How I work
Measure the pipeline, not just the kernel
Separate preprocessing, compute, data movement, synchronization, and postprocessing so local improvements can be related to end-to-end latency.
Optimize where the data points
Use tracing, sampling, counters, memory analysis, and targeted microbenchmarks to locate the limiting resource before changing code or architecture.
Keep correctness beside performance
Compare optimized paths against stable reference outputs and rerun the same measurement protocol after each material change.
Selected work
Measurement-driven machine learning
Kinship verification where evaluation changed the architecture: frozen face encoders outperformed the attempted fine-tuning path, so the simpler measured approach became the design.
Low-overhead execution tracing
C++20 scope and function tracing built around per-thread bounded buffers, Perfetto-compatible output, concurrency testing, and explicit overhead measurement.
Concurrency primitives for runtime systems
Work-stealing, dependency events, streaming statistics, false-sharing isolation, and low-level hardware helpers with explicit memory-ordering contracts and cross-platform tests.
Deterministic embedded execution
The same allocation-free C99 decision kernel runs on a phone, on microcontrollers, and in offline replay. Recorded traces, cross-architecture CI, and hardware-in-the-loop tests verify execution equivalence.
Controlled execution for AI systems
A deterministic host-owned control layer separates model proposals from authority to create effects, with independent verification and explicit policy boundaries.
Earlier
Twenty-six years at Intel included sixteen years of GPU performance tooling, low-overhead instrumentation, ETW and symbolic stack analysis, GPU memory analysis, RealSense firmware and sensor pipelines, pre-silicon firmware and OpenCL simulation, developer platforms, concurrency, computer vision, and spatial computing. That work produced eight US patents across six inventions and contribution to ECMA-363 Universal 3D.
Performance instrumentationSpatial computing architectureDetailed resume
Contact
I work on systems where performance depends on understanding both the workload and the machine: edge AI, embedded systems, computer vision, developer tooling, performance analysis, and architecture-level optimization.
Portland, OR | Berkeley, CA. Remote or hybrid. Full-time, part-time, and consulting.