PeopleofSuperintelligence

Researcher

Keshav Santhanam

Keshav Santhanam is a computer scientist who builds systems to speed up machine learning, from accelerator cluster scheduling to neural search and language-model pipelines. He earned a computer science BS at UIUC in 2017 and an MS at Stanford in 2019, and pursued a PhD at Stanford advised by Matei Zaharia, coauthoring the Gavel cluster scheduler in 2020. With Omar Khattab as joint first author, he created ColBERTv2 and the PLAID retrieval engine, both published in 2022. In 2023, he coauthored DSPy, a programming model whose compiler optimizes language-model pipelines for a chosen metric. His later work includes the ALTO orchestrator for compound AI systems in 2024, and in 2025 ColBERT-serve, the ParallelPrompt benchmark and inference work on NVIDIA's Nemotron-H models.

Badges

Timeline

  1. 2017–2019

    Degrees from UIUC and Stanford

    He earns a BS in computer science at the University of Illinois at Urbana-Champaign in 2017, working with Professor Indranil Gupta in the Distributed Protocols Research Group. He completes an MS in computer science at Stanford in 2019 with a dual concentration in systems and artificial intelligence.

  2. 2020

    The Gavel cluster scheduler

    As a Stanford PhD student advised by Matei Zaharia, he builds systems and infrastructure that accelerate machine learning workloads. In 2020, he coauthors Gavel, a heterogeneity-aware scheduler for deep learning on clusters of GPUs, TPUs and other accelerators, presented at OSDI 2020; for this work he is affiliated with Stanford University and Microsoft Research.

  3. November 2021

    The DistIR simulator

    In November 2021, he introduces DistIR with Siddharth Krishna, Ryota Tomioka, Tim Harris and Matei Zaharia. The intermediate representation and simulator identifies top-performing ways to distribute neural network computation without executing them on physical hardware.

  4. 2021–2022

    ColBERTv2 for neural search

    In December 2021, he and Omar Khattab, as equal contributors, introduce ColBERTv2, a late-interaction retriever that pairs residual compression with denoised supervision and cuts the space footprint of late-interaction models by 6 to 10 times. The paper appears at NAACL 2022.

  5. 2022

    The PLAID retrieval engine

    In May 2022, he and Omar Khattab introduce PLAID, an engine that prunes low-scoring passages through centroid interaction and cuts ColBERTv2 search latency by up to 7 times on a GPU and 45 times on a CPU. It appears at the CIKM 2022 conference.

  6. 2022–2024

    DSP and the DSPy framework

    In December 2022, he coauthors Demonstrate-Search-Predict (DSP), a framework that passes natural-language text between language models and retrieval models in multi-step pipelines. In October 2023, he joins Omar Khattab and others on DSPy, a programming model whose compiler optimizes language-model pipelines to maximize a given metric; it appears at ICLR 2024.

  7. 2024

    The ALTO network orchestrator

    In March 2024, he and Deepti Raghavan introduce ALTO, a network orchestrator for serving compound AI systems such as pipelines of language models. By streaming partial outputs between stages, it raises throughput by up to 3 times on a chatbot verification pipeline at a 4-second latency target; the work appears at the EuroMLSys 2024 workshop.

  8. 2025

    ColBERT-serve and Nemotron-H inference

    In 2025, he coauthors ColBERT-serve, accepted at ECIR 2025, which memory-maps the ColBERT index to cut RAM use by 90%. Working from Stanford University, he also coauthors ParallelPrompt, a NeurIPS 2025 benchmark of over 37,000 real-world prompts for measuring parallelism within single queries. NVIDIA's Nemotron-H hybrid Mamba-Transformer models, published in March 2025, credit him for inference work.

Career

  1. From 2020Microsoft Research

Education

  1. 2019Stanford UniversityMS in computer science
  2. 2017University of Illinois Urbana-ChampaignBS in computer science

From the biography.

  1. 16% of AI career moves in the 2020s went from a university to a companyIncludes Keshav Santhanam. We estimate that 16% of AI career moves in the 2020s went from a university to a company; 29% of moves starting at a university went to a company, and 76% ended the university role.
  2. Among employers with 30+ full-time cases, OpenAI leads primary technical role H-1B pay at $310,000, and Amazon, Google and Microsoft file 57% of all casesIncludes Microsoft. Across 440,669 certified US H-1B filings from 44 employers (FY2020–FY2026), OpenAI leads primary technical role base pay at $310,000 and xAI leads overall base pay at $290,000.
  3. Google had the most people shaping AI every year from 2014 to 2025Includes Microsoft. Our estimates of which organizations employed the most people shaping AI each year from 1955 to 2025, tracing the shift from university labs to companies.
  4. The company share of AI researchers’ top papers was 31% in 2020–2024 and 8% in 1970–1974Includes Microsoft. We estimate where the people shaping AI worked when their most-cited papers appeared: the company share was 31% in 2020–2024 and 8% in 1970–1974, and Google led all organizations in 2020–2024.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Keshav Santhanam yet. Send us a free-licensed photo.

  1. Keshav Santhanam cs.stanford.edu. Checked 20 October 2026.
  2. Heterogeneity-Aware Cluster Scheduling Policies for Deep Learning Workloads | USENIX usenix.org. Checked 20 October 2026.
  3. DistIR: An Intermediate Representation and Simulator for Efficient Neural Network Distribution arxiv.org. Checked 20 October 2026.
  4. ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction arxiv.org. Checked 20 October 2026.
  5. PLAID: An Efficient Engine for Late Interaction Retrieval arxiv.org. Checked 20 October 2026.
  6. Demonstrate-Search-Predict: Composing retrieval and language models for knowledge-intensive NLP arxiv.org. Checked 20 October 2026.
  7. DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines arxiv.org. Checked 20 October 2026.
  8. ALTO: An Efficient Network Orchestrator for Compound AI Systems arxiv.org. Checked 20 October 2026.
  9. ColBERT-serve: Efficient Multi-Stage Memory-Mapped Scoring arxiv.org. Checked 20 October 2026.
  10. PARALLELPROMPT: Extracting Parallelism from Large Language Model Queries papers.nips.cc. Checked 20 October 2026.
  11. ParallelPrompt: Extracting Parallelism from Large Language Model Queries arxiv.org. Checked 20 October 2026.
  12. Nemotron-H: A Family of Accurate, Efficient Hybrid Mamba-Transformer Models - NVIDIA ADLR research.nvidia.com. Checked 20 October 2026.