Backend engineer working on high-load asynchronous services in Python and Go. Currently focused on GPU media processing and systems programming.
trtvideo — GPU-resident TensorRT runtime for video models. Frames never leave VRAM across NVDEC → CV-CUDA → TensorRT → NVENC. Under a reproducible benchmark methodology it beats tuned VapourSynth implementations by up to 25.5% end-to-end FPS on SPAN, using 2.1–21.2× less CPU and 1.6–4.9× less peak VRAM.
safetensors.mojo —
independent implementation of the Safetensors format in Mojo. Published in the
modular-community conda channel and listed on the official Mojo package index.
Strict format validation, mmap zero-copy tensor views, deterministic atomic writes.
Linux and operating systems · parallel programming · GPU computing · compilers · video processing (codecs, encode/decode, colour conversion) · Web3