Arnur Yembergen

ML systems engineer: GPU inference, low-precision formats, systems programming.

coldblime.com · github.com/1morello

experience

ML engineer, inference · [company]

[2024 - present]

[what got faster, by how much, on which hardware]
[what you built or owned to get there]

[earlier role] · [company]

[2021 - 2024]

[scope of the work]
[one result worth naming]

education

[degree] · [university]

[years]

skills

CUDATritonvLLMPyTorchNVFP4 / quantization

selected results

fp4-speech. Hardware FP4 inference for speech on Blackwell: FP4 weights safe across three architectures, 2x Orpheus decode with CUDA graphs, activation failure modes measured and mapped.

in progress