Aurko Roy
Senior Research Scientist at Google
San Francisco, California, United States
Summary
👤
Senior🎓
Top SchoolAurko Roy is a Senior Research Scientist at Google with 11 years of experience applying optimization, generative models, and NLP to large-scale, production problems. He holds a PhD in Algorithms, Combinatorics & Optimization from Georgia Tech and has driven impactful research-to-product work such as Routing Transformer deployments across YouTube Search, Google Search, Docs and Gmail that improved latency and model quality. At Google he open-sourced Lingvo JAX, helped train a 200B-parameter model for MLPerf, and publishes on latent n-grams and transformer improvements, blending systems engineering with theoretical rigor. His open-source contributions include implementing adversarial robustness features in CleverHans and model optimizations in Google Research and tensor2tensor, revealing a hands-on approach to both research code and production-grade model engineering. Colocated in San Francisco, he’s interested in applied research roles in tech and quantitative finance where optimization meets large-scale language modeling.
11 years of coding experience
4 years of employment as a software developer
Indian Institute of Technology KanpurDoctor of Philosophy (PhD), Algorithms, Combinatorics and Optimization (ACO), 4/4, Doctor of Philosophy (PhD), Algorithms, Combinatorics and Optimization (ACO), 4/4 at Georgia Institute of Technology
Chinese, English
Stackoverflow
1reputation
0reached
0answers
0questions
Github Skills (21)
python10
attention-mechanism10
machine-learning10
autoencoder10
vector10
transformer-models10
keras10
deep-learning10
tensorflow10
quantization10
image-processing9
benchmark9
security9
benchmarking9
nlp9
Programming languages (2)
Github contributions (5)
tensorflow/tensor2tensor
Mar 2018 - Jun 2020
Library of deep learning models and datasets designed to make deep learning more accessible and accelerate ML research.
Role in this project:
ML Engineer
ML EngineerContributions:47 commits in 2 years 3 months
Contributions summary:Aurko primarily worked on the `tensor2tensor` library, contributing to the implementation and refinement of discrete vector quantization (DVQ) techniques. They made several changes to the `discretization.py` file, including updates to the `embedding_lookup` and `nearest_neighbor` functions, and the addition of a test for transformer_vae. Their contributions appear focused on improving the performance and stability of DVQ within the context of autoencoders and transformer models. Further work on the model involved adjusting the decoder to use non-causal attention and updating code to adapt to new APIs.
pytorchautoencoderdeep-learningmachine-translationreinforcement-learning
google-research/google-research
Oct 2020 - Aug 2021
Google Research
Role in this project:ML Engineer
ML EngineerContributions:30 commits in 10 months
Contributions summary:Aurko primarily contributed to the `google-research/google-research` repository by implementing and modifying components related to the Routing Transformer architecture. Their work involved the creation and modification of the `sparse_image_transformer.py` file, suggesting a focus on image-related problems. The user integrated GEGLU activation, optimized attention mechanisms using einsums, and added a TF inference API, demonstrating both model implementation and optimization skills. Furthermore, the user updated hyperparameters and README files, suggesting participation in model configuration and documentation.
googlemachine-learningai
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.