Hanzi Mao
Staff Research Scientist at Google DeepMind
Mountain View, California, United States
Summary
🤩
Rockstar🎓
Top SchoolHanzi Mao is a Staff Research Scientist with 10+ years building foundation models for vision, language, and embodied AI, currently working on Gemini Robotics at Google DeepMind. Previously he led research and engineering for NVIDIA’s Deep Imagination Research lab, where he was the technical lead on Cosmos-Predict2 and Cosmos-Transfer1 and drove diffusion-based pretraining for the Cosmos family announced at CES 2025. Before that he was a core contributor at Meta AI to high-impact projects like Segment Anything, ViTDet and ConvNeXt, and has hands-on experience extending Detectron2 and PyText with practical features such as federated loss, Swin backbone support, and two-tower decoders. He holds a PhD in Computer Science from Texas A&M and combines deep academic training with product-focused engineering—having shipped multimodal models into multiple products and downstream systems. Colleagues describe him as both a systems-minded researcher who builds open, adaptable world models and a pragmatic engineer who ties state-of-the-art research to reproducible tooling.
10 years of coding experience
10 years of employment as a software developer
Doctor of Philosophy (PhD) Computer Science, Doctor of Philosophy (PhD) Computer Science at Texas A&M UniversityMaster of Engineering (M.Eng.) Electronics and Telecommunications Engineering, Master of Engineering (M.Eng.) Electronics and Telecommunications Engineering at Huazhong University of Science and Technology
Github Skills (17)
pytorch10
python10
semantic-segmentation10
machine-learning10
trainings10
natural-language-processing10
roberta10
object-detection10
computer-vision10
nlp10
modeling10
quants9
federated-learning9
sd-mmc9
swin-transformer9
Programming languages (2)
Github contributions (5)
facebookresearch/pytext
Aug 2020 - Jul 2021
A natural language modeling framework based on PyTorch
Role in this project:
ML Engineer
ML EngineerContributions:8 commits, 5 PRs in 11 months
Contributions summary:Hanzi made significant contributions to a two-tower document classification model within the PyText framework. They implemented a new two-tower MLP decoder and a corresponding TwoTowerClassificationModel, enhancing the framework's flexibility. Further contributions included supporting text inputs from both sides of the model and adding support for graph mode quantization, specifically for Linformer encoders. They also worked on exporting embeddings, loading pretrained model states for the MLP decoder, and introduced a shared encoder option.
pytorchnlpbertmachine-learningnatural-language
facebookresearch/ConvNeXt
Jan 2022 - Jan 2022
Code release for ConvNeXt model
Role in this project:ML Engineer
ML EngineerContributions:7 commits, 7 pushes, 24 comments in 6 days
Contributions summary:Hanzi primarily focused on configuring and updating model training pipelines. This included adding license headers to configuration files and differentiating between single-scale and multi-scale configurations for semantic segmentation models. The commits also involve the modification of backbone and model initialization files, suggesting a hands-on approach to model definition and training setup. Furthermore, the user added a customized text logger for training metrics.
imagenetconvnextvisual-recognitionbackbone
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.