Afroz Mohiuddin
Member Of Technical Staff at OpenAI
Mountain View, California, United States
Summary
🤩
Rockstar🎓
Top SchoolAfroz Mohiuddin is a Member of Technical Staff with 14 years of experience building large-scale distributed systems and core deep learning research, currently contributing to OpenAI after senior research engineering roles at Meta and Google. He has driven projects across Search, LLMs and generative models (Gemini, MUM, Google Brain) and brings hands-on expertise in production ML pipelines and model training optimizations. An active open-source contributor, Afroz has improved influential repositories such as T5/T5X and Trax—adding novel metrics, performance logging, and tooling that make large-model evaluation and training more robust and observable. He combines research rigor from IIT Kanpur with production sensibilities, often focusing on subtle but impactful infrastructure details like effective batch-size logging, pre-emption-aware uptime metrics, and sklearn metric wrappers. Colleagues rely on him for bridging cutting-edge research ideas into reliable, scalable systems used at internet scale.
14 years of coding experience
15 years of employment as a software developer
Indian Institute of Technology Kanpur
English, Hindi, Urdu
Github Skills (37)
data-pipelines10
python10
evaluation10
data-set10
scikit10
machine-learning10
filetype10
file-conversion10
file-operations10
data-management10
file-format10
data-model10
datasets10
trainings10
tensorflow10
Programming languages (4)
Github contributions (5)
google/trax
Oct 2019 - Oct 2021
Trax — Deep Learning with Clear Code and Speed
Role in this project:
ML Engineer
ML EngineerContributions:16 releases, 14 reviews, 258 commits in 2 years 1 month
Contributions summary:Afroz made several changes to the `trax` repository, primarily related to machine learning tasks. The commits focused on defining new learning rate schedules and also made changes to include the T5 text pre-processing code. Additionally, the user added and modified configurations for testing and fine-tuning models, including those used for reinforcement learning in an Atari environment, and implemented functionalities like text dataset batching with span corruption.
pytorchspeeddeep-learningnumpyreinforcement-learning
google-research/text-to-text-transfer-transformer
Feb 2020 - Mar 2021
Code for the paper "Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer"
Role in this project:ML Engineer
ML EngineerContributions:14 commits in 1 year 1 month
Contributions summary:Afroz primarily focused on enhancing the T5 model, contributing new metrics and functionalities. They added an AUC metric, including an option to threshold targets for regression tasks. Furthermore, the user implemented a wrapper for sklearn metrics, enabling the integration of various metrics from the sklearn library. They also introduced a feature for sampling a batch of evaluation examples for large datasets and made several adjustments to the code to reduce log spam.
pytorchnlplimitsberttransfer-learning
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.