Jacob Devlin
Seattle, Washington, United States
Summary
🤩
RockstarJacob Devlin is a software engineer with seven years of experience building scalable systems at major tech firms, currently working at Google after three years at Microsoft. Based in Redmond, he blends production engineering with applied machine learning, notably contributing to the widely used google-research/bert project by improving SQuAD tooling, reducing CPU memory use via TFRecord pipelines, and adding SQuAD 2.0 and whole-word masking support. His work reflects a pragmatic focus on efficient data handling for large NLP models and making research code more production-ready. With a Master's from the University of Maryland, he brings both academic grounding and hands-on implementation skills. Colleagues rely on him to bridge research ideas and engineering constraints, turning model improvements into robust, deployable components.
7 years of coding experience
Github Skills (9)
pytorch10
machine-learning10
nlp10
tensorflow10
bert10
python9
modeling8
natural-language-processing8
google7
Github contributions (3)
google-research/bert
Oct 2018 - Mar 2020
Role in this project:
ML Engineer
ML EngineerContributions:65 commits, 39 PRs, 56 pushes in 1 year 4 months
Contributions summary:Jacob's commits primarily focus on modifying and enhancing the `run_squad.py` file, which is responsible for running BERT on the SQuAD dataset. They updated the code to reduce CPU memory usage by writing a TFRecord file, demonstrating an understanding of efficient data handling for large datasets. Additional commits include adding support for SQuAD 2.0 and incorporating whole word masking techniques to improve model performance.
nlpnatural-language-understandingbertgooglepre-trained
paperplanet/bert_finetune_reproduce
Oct 2018 - Mar 2019
Contributions:60 commits in 4 months
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.