My team led and built end-to-end TPU system support for Google products, especially super large embedding framework for recommendation models.
I also drove the innovative design, which introduces the distributed tensor representation on top of TensorFlow 2.0. This is an umbrella project to enable within-layer model parallelism, pipelining, and offloading. It powers large models like GPT-3--new hotness in cutting-edge machine learning.
I worked with lots of great people to define and develop TensorFlow Hign-Level machine learning framework; see Google Research Publication.
My area of technical expertise is CPU/TPU distributed training, especially in super large clusters. My work directly empowered countless Google products, research projects like AlphaZero, production platforms like Cloud TPU and TFX, etc.
Owner of multiple critical services, including Songza-style Concierge, Configuration, Notification, Play Store Integration, Radio Station Landing Page, Concerts, Unified Sync, etc.
20% Project: I contributed to machine learning based Music recommendation system to improve cold-start problem so new users' experience could be better.