Viraj Agrawal

Research Fellow
Microsoft Research India
Bengaluru, India

virajagr [at] gmail [dot] com
t-viragrawal [at] microsoft [dot] com

CV · Google Scholar · ORCID · LinkedIn

I am a Research Fellow at Microsoft Research India, where I work on algorithms for retrieval: how to search large collections of vectors efficiently, and what vector representations can and cannot express.

Much of my work is on graph-based nearest neighbor search, the approach behind DiskANN. With Kirankumar Shiragur, I proved approximation guarantees for adaptive beam search on α-reachable graphs. We then used the analysis to design a learned beam-width predictor that reduces query latency at the same recall (under review). With Ravishankar Krishnaswamy and Magdalen Dobson Manohar, I am now working on improving search in Filtered-DiskANN, where each query carries attribute filters alongside its vector.

I have also worked on multi-vector retrieval with Kirankumar Shiragur and Ankit Garg. We built a family of retrieval tasks where ranking the relevant documents above the irrelevant ones requires exponentially large single-vector embeddings, while polynomial-size multi-vector embeddings suffice, and turned the construction into a benchmark, ANDOR. With Ankit, I am now working on approximating multi-vector representations with single vectors for first-stage retrieval.

Before joining Microsoft Research, I was a quantitative researcher at Quantbox Research for a year, looking for short-horizon predictive signals in Indian equity and index-options markets. Competitive programming was a serious hobby for about four years: I represented IIT Delhi at the ICPC Asia West Finals 2024, and eventually spent as much time setting and testing problems as solving them.

I am open to PhD positions and research roles in industry and academia.

Earlier