Justin Zhao

Research Engineer @ Meta

About

AI researcher at Meta Superintelligence Labs with a focus on evaluations for trust and safety.Before Meta, I was at Google Research (2015-2021) where I worked on the first text generation model for the Google Assistant. I left to build an enterprise platform for LLM post-training as part of a startup that eventually got acquired. During some time off 2024-2025, I became interested in LLM judges and published the Language Model Council, a democratic twist on LLM evals, and have been focused on evaluation methods ever since.

Country

United States

City

San Francisco Bay Area

Industry

Computer Software

Skill

Technical Leadership, Management, TensorFlow, Linguistics, Natural Language Processing (NLP), Open-Source Software, PyTorch, Distributed Systems, Python, Deep Learning, Tensorflow, Composition, Artificial Intelligence, Machine Learning, Java, JavaScript, tensorflow, C/C++, Unix, Adobe Creative Suite

Experience

Research Engineer

Technical Advisor

Locate

2024-11 - Present · 1 yr 11 mos

- ML for revenue prediction for retail businesses. - Brokerage automation using agents + MCP.

Staff Software Engineer, Founding

Predibase (acquired by Rubrik)

2022-10 - 2024-8 · 1 yr 11 mos

TL(M) of the core Machine Learning team, later transitioning to an IC researcher role, focusing on efficient LLM post-training.

Founding Engineer

Predibase (acquired by Rubrik)

2021-10 - 2024-8 · 2 yrs 11 mos

New York, United States

- Ludwig (https://github.com/ludwig-ai) - Horovod (https://github.com/horovod/horovod) - Ray (https://github.com/ray-project/ray)

Senior Research Engineer

Google

2015-9 - 2021-9 · 6 yrs 1 mo

Generative language models on Google Search and the Assistant. Search and Assistant, Mountain View (2017-2021) - Continuous ML infrastructure for Transformer-LSTM model distillation from synthetic data generation to deployment, 30+ languages, 100’s of models / week - mostly C++/Tensorflow/lingvo. - Evaluation algorithms for benchmarking language models. - On-device tf.Lite text generation models (~8MB compiled) for the Google Home Mini. - 3X presenter at Google Research Conference, hosted 1 intern (+conversion), 60+ interviews, PhD Panelist at Stanford & Columbia. Research and Machine Intelligence, New York (2015-2017) - Custom multi-layer LSTMs applied to academic NLG datasets (E2E, WebNLG, MultiWOZ). - Linguistically intelligent template-based system (100’s clients, 500k templates, 50+ languages). - Text generation for weather in English with public alerts, ~4 million queries per day. - Large-scale refactoring and API migration efforts (~1K Java files), affecting ~700M QPD on Search infrastructure (GWS).