Piotr Żelasko
Principal Research Scientist at NVIDIA
Miami, Florida, United States
Summary
🤩
Rockstar🎓
Top SchoolPiotr Żelasko is a Principal Research Scientist at NVIDIA with a decade of experience building and optimizing automatic speech recognition and audio processing systems across academia and industry. He holds a PhD in Electronics Engineering and has led research teams (Meaning) and contributed to leading ML toolchains at JHU, Avaya, and core open-source projects like Lhotse and Icefall, improving data handling, feature extraction, and CUDA memory strategies for scalable training. Piotr combines deep acoustic and ML expertise with hands-on engineering—C++, Python, and CI/dockerized deployments—to take prototypes into production and prevent common failure modes such as OOMs. He has practical experience in telephony ASR, IVR integrations, and punctuation-restoration NLP for cleaner transcripts, reflecting a focus on real-time, robust systems. Based in Miami, he balances rigorous research with creative outlets—playing blues and jazz trombone—which hints at his intuitive grasp of audio beyond algorithms.
10 years of coding experience
9 years of employment as a software developer
AGH UniversityMaster’s Degree, Acoustic Engineering, Master’s Degree, Acoustic Engineering at AGH University of Krakow
Polish, English
Stackoverflow
1reputation
0reached
0answers
0questions
Github Skills (19)
pytorch10
data-pipelines10
python10
feature-extraction10
machine-learning10
audio-processing10
data-processing10
cuda10
speech-recognition10
data-pipeline10
data-science9
webdataset9
deep-learning9
kaldi8
json8
Programming languages (10)
Github contributions (5)
lhotse-speech/lhotse
Apr 2020 - Jan 2023
Tools for handling speech data in machine learning projects.
Role in this project:
Backend Developer & ML Engineer
Backend Developer & ML EngineerContributions:42 releases, 733 reviews, 1958 commits in 2 years 9 months
Contributions summary:Piotr implemented and refactored features related to data processing and feature extraction within the Lhotse speech processing library. The commits demonstrate work on adding new functionality for lazy manifest handling, including methods to manipulate and manipulate cuts, improving both the flexibility and memory efficiency of the library. The commits also integrate and develop new audio and features extractors, alongside performance optimizations.
pytorchkaldipythondatamachine-learning-projects
k2-fsa/icefall
Oct 2021 - Jan 2022
Role in this project:ML Engineer
ML EngineerContributions:90 reviews, 22 commits, 11 PRs in 3 months
Contributions summary:Piotr primarily contributed to the training and data processing aspects of the "icefall" project, a speech processing repository using k2 and Lhotse. Their work included implementing data sampling strategies, integrating CUDA memory optimization, and refactoring code related to training pipelines. Key contributions involved modifying training scripts to find and prevent CUDA out-of-memory errors before starting the training loop.
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.