Raul Puri
United States, United Kingdom
Summary
🤩
RockstarRaul Puri is a researcher and machine learning engineer with 10 years of experience based in Berkeley, currently on OpenAI’s Multimodal Learning team working on GPT-4 multimodality, Codex, embeddings, and the GPT Edit feature. He brings research leadership from NVIDIA and applied ML roles, spanning unsupervised and curriculum learning, distributed and mixed-precision training, and production deployment of speech and NLP systems. An active open-source contributor, he has helped adapt high-profile NVIDIA projects—patching Tacotron2 for reliable single-GPU and distributed inference and scaling unsupervised language modeling for robust sentiment classification—down to CUDA fixes and even changing optimizer defaults to improve training stability. With an EECS degree and near-complete bioengineering minor from UC Berkeley, he blends systems-level engineering with algorithmic research and a practical focus on reproducible ML at scale.
11 years of coding experience
Github Skills (24)
pytorch10
distributed-training10
python10
machine-learning10
speech-synthesis10
deep-learning10
cuda10
nlp10
preprocess9
datapreprocessing9
preprocessing9
pre-processing9
data-prep9
model-optimization9
data-pre-processing9
Programming languages (4)
Github contributions (5)
NVIDIA/sentiment-discovery
Dec 2017 - Oct 2018
Unsupervised Language Modeling at scale for robust sentiment classification
Role in this project:
ML Engineer
ML EngineerContributions:3 releases, 27 commits, 9 PRs in 10 months
Contributions summary:Raul primarily contributed to the development and maintenance of the sentiment discovery model, addressing critical issues such as import changes, CUDA compatibility, and handling of different data types. They fixed bugs, updated data loading procedures, and integrated necessary dependencies. Furthermore, the user modified the model wrapper and main script by changing default optimizer to Adam and added a base gpu argument for distributed training.
pytorchnlpbertdeep-learningunsupervised
NVIDIA/tacotron2
May 2018 - May 2018
Tacotron 2 - PyTorch implementation with faster-than-realtime inference
Role in this project:ML Engineer
ML EngineerContributions:7 commits, 3 PRs, 6 pushes in 2 days
Contributions summary:Raul primarily focused on updating and adapting the Tacotron2 model for a new version (0.4), including modifications to training scripts, model definitions, and utility functions. They adjusted the model's data handling, particularly regarding input lengths and padding. Additionally, the user patched the inference script to address issues with distributed data parallel models and single GPU execution.
pytorchrealtimeinferencefastertacotron
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.