I'm a Senior Research Scientist at Databricks Mosaic, NYC, where I build high-performance enterprise AI agents, including KARL. My research focuses on using synthetic data and reinforcement learning to train language-model-powered agents.

Previously, I worked at NVIDIA on the Nemotron model family and the OpenMath series, and at FAIR, Meta AI. I received my Ph.D. from TTI Chicago, advised by Kevin Gimpel and Karen Livescu, and my B.Tech. in Computer Science from IIT Kanpur.

My research focuses on training language-model agents and understanding how they represent, search, and reason over information. My work spans two connected areas:

Synthetic Data for Post-Training ¶ I develop synthetic-data recipes for training and evaluating state-of-the-art knowledge agents. At Databricks, this work has powered the training of models such as KARL and Instructed-Retriever-1, and was central to developing OfficeQA Pro V2, a benchmark for enterprise grounded reasoning. Previously at NVIDIA, I was a primary contributor to the OpenMath series and also contributed to the Nemotron model family. This work, including OpenMathReasoning, helped our team win the AI Math Olympiad 2.

State Tracking and Memory ¶ I study how language models represent and update world state across long contexts. My work has demonstrated how latent state tracking emerges in language models trained on chess games and shown the benefits of code pretraining for entity tracking. I also introduced Self-Notes, which lets models externalize intermediate reasoning and state information on the fly. During my PhD, I developed state-of-the-art memory-based models for people tracking and long-document coreference. More recently, Major Entity Identification and IdentifyMe extended this line of work to modern language models and long-context evaluation.

1. Aug 2026

   OfficeQA Pro V2 establishes a new benchmark for enterprise grounded reasoning.

2. Jun 2026

   Instructed-Retriever-1 uses parallel test-time scaling to make search over 3× faster.

3. May 2026

   MemEx equips LLM agents with a programmable scratchpad for using code as action.