Mosaic AI research @ Databricks

VLMs, RL, MoEs, and LoRAs

https://www.databricks.com/blog/pdfs-production-announcing-state-art-document-intelligence-databricks

Lead for open and closed loop model development and various model platform projects

• Optimized the training efficiency of reinforcement learning models and expanded its training capacity

• Unblocked the deployment of reinforcement models

• Enabled researchers to design and experiment interactive and RL models

• Accelerated data processing pipelines that uses model inferencing

https://medium.com/nuro/jonathan-booher-wei-liu-zixun-zhang-joyce-huang-ethan-tang-18db008d5d08

Worked on designing foundational platforms for self driving model development

• Lead the migration effort of models from Tensorflow 1 to Tensorflow 2 + Keras

• Optimized training efficiency of various perception models to cutdown training time and save cost

• Unblocked various model development and optimization efforts such as QAT and pruning

• Developed general model training libraries to accelerate model development

Worked on Nvidia's autonomous vehicle infrastructure team (Maglev) to bring NVAutonet into production

• Prototyped model development in PyTorch

• Optimized model training performance to train faster

• Implemented better loss balancing algorithms to improve model accuracy