Tushar is an Applied Scientist-2 at Amazon Search in the M5 team, working on building representations that are trained on 1000x more data, & have 1000x more parameters than the largest models in current use at Amazon. He helped create index-able & searchable images embeddings along with text using multimodal models like CLIP & ALBEF. This helps improve the suite of tasks & is adopted by Search, Ads CTR team - leading to predicted lift of $48.7MM OPS. Prior to this, he worked as Research Engineer & then Research Scientist at Verisk Analytics, on multiple projects like using active learning to speed up training of LLMs by 2.5x, Document Intelligence tasks (ACL 2023), & vector-quantized normalizing flows (UAI 2022). Previously, at NYU he worked with esteemed professors like Yann Lecun on developing world models RL agents for text-based games via memory-based NN. Moreover, with Prof Rob Fergus, he developed novel architecture for multi-agent communication in RL (ICLR 2019).