Sheshansh Agrawal
Director of Research at Contextual AI. I work on making retrieval and agentic search systems more accurate, faster, and generalizable. From state-of-the-art rerankers to structured data retrieval to agent and tool design.
Before this, I was at Microsoft Research, where I worked with Prof. Manik Varma on extreme classification, personalized dense retrieval algorithms, and Bing Ads where I built a scalable training method and infrastructure for embeddings. My models and algorithms served 100s of millions of users daily, and generated O($X00M) in revenue.
Before that, I did my Bachelors in Computer Science from IIT Bombay, where I did my thesis with Prof. Soumen Chakrabarti.
Research
Agentic Search
Incubated the Agent Composer product at Contextual, working on agent orchestration, particularly dynamic agent loops and tool integration. Built an agentic alternative to GraphRAG, demonstrating the power of careful tool design in agentic search. Also worked on parametric optimization (RL) and non-parametric optimization (prompt optimization). Currently working on and thinking about memory and continual learning.
Rerankers
Built state-of-the-art pointwise rerankers, including the world's first instruction-following reranker. Published BlitzRank, a Pareto-optimal LLM ranking algorithm that uses 25–40% fewer tokens than comparable methods at near-identical quality.
Retrieval
Built CORGEE, a scalable training method and infrastructure for embeddings. Scaled to trillions of tokens across hundreds of GPUs, serving over a quarter of retail ads on Bing globally. Developed XPERT, the first personalized dense retrieval algorithm at scale (SIGIR 2023). Extended extreme classification to handle richer metadata with DECAF (WSDM 2021) and ECLARE (WWW 2021). Improved out-of-domain generalization for approximate nearest neighbor search with OOD-DiskANN.
Structured Retrieval
Built a Text-to-SQL system that topped the BIRD-Bench leaderboard, demonstrating that local models can compete with closed-source giants on enterprise-scale structured data tasks.