Sharan Narang
Research Engineer, Facebook AI at Meta
San Francisco, California, United States
Summary
🤩
Rockstar🎓
Top SchoolSharan Narang is a senior AI research leader with 11 years of experience building large-scale foundation models and ML infrastructure, currently leading AI Foundations at Waymo in San Francisco. Previously he directed Llama pre‑training at Meta, shaping strategy and technical vision for Llama 2–4 and pioneering MoE, ultra‑long contexts, distillation, and synthetic data pipelines. His earlier career includes technical leadership on PaLM and T5 at Google and deep learning systems and benchmarks at Baidu, reflecting a rare blend of model research, distributed training, and systems engineering. A prolific open-source contributor, he has helped advance Mesh TensorFlow, TFDS, multilingual T5 and other high‑impact Google research repos—work that underpins modern model parallelism and dataset tooling. Known for translating research prototypes into production at hyperscale, he pairs hardware-aware optimization experience from NVIDIA and AMD with academic training from Georgia Tech.
11 years of coding experience
Masters of Science, Masters of Science at Georgia Institute of TechnologyBachelors of Engineering, Bachelors of Engineering at University of Mumbai
Github Skills (22)
multilingual10
parallelization10
python10
data-set10
machine-learning10
tpu10
datasets10
data-preprocessing10
transformer-models10
tensorflow10
natural-language-processing10
nlp10
distributed-training9
data-model9
data-management9
Programming languages (3)
Github contributions (5)
google-research/multilingual-t5
Feb 2021 - Sep 2022
Role in this project:
ML Engineer
ML EngineerContributions:7 commits in 1 year 7 months
Contributions summary:Sharan primarily focused on modifying the task definitions within the multilingual-t5 project. They updated dataset versions for tasks like TyDiQA, incorporated the "mt5" prefix into NER tasks, and introduced helper functions to manage XNLI and XQuAD tasks. Their contributions also involved adjustments to pre-training tasks and mixtures.
google-research/text-to-text-transfer-transformer
Nov 2019 - Feb 2022
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:32 commits, 25 comments, 14 issues in 2 years 3 months
Contributions summary:Sharan made several contributions related to the T5 text-to-text transfer transformer model. They fixed issues within the `transform_checkpoints` function, ensuring proper operation in different modes. Further modifications included updates to the model implementation for Mesh-TF, including finetuning capabilities, and the addition of a scoring mechanism. These changes suggest a focus on improving model functionality and performance within the T5 framework.
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.