Hey there! I’m Tushar, and I’m currently working as Sr Applied Scientist at Amazon Search with the M5 team. My main focus is on building foundational models like Amazon Titan and serving LLMs at scale through AWS Bedrock. Recently, I’ve also worked on developing multimodal LLMs and embeddings that improve search, ads, and CTR at Amazon—some of these efforts are predicted to drive a lift of $48.7MM in OPS!

Before Amazon, I spent time as a Research Engineer and then Research Scientist at Verisk Analytics, where I tackled exciting challenges like using active learning to accelerate LLM training by 2.5x, working on Document Intelligence tasks (ACL 2023), and contributing to vector-quantized normalizing flows (UAI 2022).

I also have a research background from NYU, where I worked with professors like Yann LeCun on reinforcement learning agents for text-based games, and Rob Fergus on multi-agent communication in reinforcement learning (ICLR 2019).

Looking forward to connecting and learning together!