Skip to content

Passion

Projects

I work on whatever feels most interesting and rewarding at the time.

All repositories

10 of 10 projects

HTTP/3 for httpd on OpenSSL 3.5 QUIC and nghttp3: 6,409 lines of C across 39 files, 61 signed releases, 56 integration tests, seven CI workflows, and an upstream change to httpd's process management.

  • C
  • QUIC
  • OpenSSL 3.5
  • nghttp3
  • Apache httpd

Do a protein language model and a chemical language model say different things about a peptide? Four fusion designs, four properties, one public benchmark: no. Test F1 within 0.03 of the state of the art, and a negative result that narrows the search.

  • JAX
  • Flax NNX
  • Grain
  • Cloud TPU
  • ESM-2

A small typed library that makes the hidden choices at JAX ecosystem seams explicit: array ownership, trace budgets, key derivation, checkpoints, sharding, weight loading. On PyPI, with a nine-way CI matrix.

  • Python
  • JAX
  • Flax NNX
  • Optax
  • Orbax

The CMake layout I start C++ and GPU projects from: toolchain selection, sanitizers, ccache, clang-format and clang-tidy, CPM-managed dependencies, and a CUDA or HIP runtime chosen at configure time.

  • C++20
  • CMake
  • CUDA
  • HIP
  • CPM.cmake

MuZero from scratch: representation, dynamics, and prediction networks with Monte Carlo tree search, on Gymnasium environments.

  • Python
  • JAX
  • Flax NNX
  • Gymnasium
  • Cloud TPU

Seven small repositories from the operating systems and real-time courses: system-call and interrupt simulation, process and scheduler simulation with four policies, IPC over System V shared memory and semaphores, bare-metal MSP432 examples, and a readers-writers solution in C++.

  • C
  • POSIX
  • System V IPC
  • Semaphores
  • MSP432

Vision transformer training on Cloud TPU pods (1 TB HBM, 10 petaFLOPs FP16): mesh topology, per-host utilization, and honest throughput numbers. NeurIPS 2025.

  • JAX
  • Cloud TPU
  • Vision Transformers
  • Python