• Oct 2026: New preprint on machine unlearning for federated socially fair k-means.
• Sep 2026: X-AVDD, on audio-visual dataset distillation, was accepted at NeurIPS 2026.
• Sep 2026: Google awarded CeRAI a US$100,000 research gift to support my work on the multilingual and multicultural safety of agentic AI systems.
• Jul 2026: Released Pluralis v0.1, a multicultural, multimodal, multilingual benchmark for AI risk and reliability.
• May 2026: Recognized as a Silver Reviewer for the ICML 2026 Position Paper Track.
• Nov 2025: DiRe, on dataset condensation, was accepted at WACV 2026.
• Nov 2025: Joined the Centre for Responsible AI at IIT Madras as a Principal Project Scientist.
• Apr 2025: Granted a US patent on online inference for nonsymmetric DPPs.
• Jul 2024: Joined the Department of AI at IIT Hyderabad as a Visiting Assistant Professor.

My research is in Theoretical Machine Learning, grounded in theoretical computer science, and extends into Responsible AI and Computer Vision. Theory is the foundation: my work in the other two areas builds on it.

Theoretical Machine Learning

I design algorithms with provable guarantees, often going beyond worst-case analysis to explain why algorithms work on real data even when their worst-case bounds are pessimistic. My three best papers are from this line of work:

• One-pass MAP inference for nonsymmetric DPPs (ICML 2022), which started the study of streaming and online MAP inference for these models. I am the sole first and corresponding author, and the work led to a granted US patent.
• LP recovery guarantees for MAP inference (AISTATS 2021), which gives a new explanation of why a standard LP relaxation succeeds on real vision instances. I am co-first author.
• Improved guarantees for k-means++ and its parallel variant (NeurIPS 2020), which improves the k-means++ bound from 8(ln k + 2) to 5(ln k + 2) for an algorithm that is the default in widely used libraries.

I have also worked on dynamic and adaptive algorithms for tensor product regression (NeurIPS 2022), pairwise summation (IEEE BigData 2022) and Mahalanobis distance estimation (IEEE BigData 2023).

Responsible AI

I use these tools to ask when an AI system can be trusted: whether it is fair, whether it can forget data on request, and whether it is safe in the cultural contexts where it is used.

• With Dr. Sunayana Sitaram, I am India lead of Pluralis, a culture-first multimodal safety benchmark that is part of the MLCommons AILuminate effort, with 6,448 image-text prompts across six Asia-Pacific locales and eight languages. CeRAI led benchmark design, curation and validation for English, Hindi and Tamil.
• Under a Google Research Gift, I lead CeRAI’s effort on multilingual, multicultural evaluation of AI agents.
• With Prof. Saurav Prakash, I work on fair, deletable clustering. We prove an O(log² k) guarantee for federated socially fair k-means that does not depend on how groups are split across institutions, using our NeurIPS 2020 bound as a black box, and we show that replay reproduces retraining exactly.

Computer Vision

Training vision models for genomics at the Broad Institute took months of compute, which led me to dataset distillation: synthesizing a small dataset with nearly the training value of a large one. With Dr. Konda Reddy Mopuri and Saumyaranjan Mohanty, I developed DiRe (WACV 2026), a diversity regularizer that draws on my work on diverse subsets and improves eight leading methods from CIFAR-10 to ImageNet-1K, and X-AVDD (NeurIPS 2026), which brings distillation to audio-visual data, cutting synthesis time from 10 to 1.3 GPU-hours and roughly tripling accuracy on VGGSound.

Research plan

In the next two to three years, I want to build a group of PhD and MS students, funded initially by the Google gift (until March 2028), around three directions that apply theory to open problems from my recent work:

1. Fair and frugal multilingual evaluation: compressing AI safety benchmarks into small representative subsets while minimizing the worst error across languages, so that low-resource languages are not sacrificed for accuracy on English.
2. Diversity-aware benchmark creation: treating benchmark construction as online diverse-subset selection, so that each new item is accepted only if it adds coverage.
3. Diversity with guarantees for data-efficient vision: replacing heuristic diversity penalties with principled objectives, and proving when a distilled dataset preserves the training signal of the full data.

In the longer term, I aim for a theory of how little data, how few evaluations and how cheap an update an AI system needs while remaining fair and safe.

Research funding

Google Research Gift, US$100,000 (awarded Sep 2026; project Oct 2026 to Mar 2028)
Scaling Pluralis to Enable Multilingual AI Safety Evaluation for Agents Across the Globe.
Co-authored the proposal with Dr. Lora Aroyo (Google DeepMind); project lead at CeRAI. PI: Prof. Balaraman Ravindran, IIT Madras.

Adobe Research Gift, US$20,000 (2022)
Dimension Reduction to Enable Learning on the Edge.
Co-authored the proposal with Dr. Zhao Song and Dr. Ritwik Sinha (Adobe Research). PI: Prof. Aravindan Vijayaraghavan, Northwestern University.

Selected awards

• Silver Reviewer Award, ICML 2026 Position Paper Track (2026)
• NeurIPS Scholar Award, US$2,000 (2022)
• Peter and Adrienne Barris Outstanding TA Award, Northwestern University (press release) (2022)
• CS Supplemental Award, US$3,000, Northwestern University, for top CS Ph.D. applicants (2018)
• Charpak Scholarship for Research Internship, Government of France (2017)
• All India Rank 189, JEE Advanced (2014)
• Gold Medallist, Indian National Physics Olympiad (2014)

I taught four courses as a Visiting Assistant Professor at IIT Hyderabad and have given guest lectures in Machine Learning I at IIT Madras.

Instructor, IIT Hyderabad

Ratings are out of 5.

Teaching interests

I prefer courses of a mathematical nature related to machine learning and algorithms, and I can teach core courses in both CSE and Data Science and AI, such as Data Structures and Algorithms, Discrete Mathematics, Machine Learning, Optimization, and Probability and Statistics. I would also like to develop two new courses:

• Beyond Worst-Case Analysis, with a focus on topics such as clustering, robust statistics, topic models and algorithms with predictions.
• Submodular Optimization and Applications to Machine Learning, a full 3-credit extension of the course I taught at IIT Hyderabad.

Guest lectures

• Maximum Likelihood and Probabilistic Linear Regression, Machine Learning I, IIT Madras (Aug 2026; one week, three lectures)
• Submodular Function Maximization, Algorithms for Big Data, Northwestern University (Apr 2022)
• Introduction to Markov Chains, Math Toolkit for TCS, Northwestern University (May 2019)

Teaching assistant

• Mathematical Foundations of Computer Science (CS 212), Northwestern University: Spring 2022, Fall 2020, Fall 2019. Received the Peter and Adrienne Barris Outstanding TA Award for Spring 2022.
• Design and Analysis of Algorithms (CS 336), Northwestern University: Winter 2019
• Fundamentals of Computing (ESC 101), IIT Kanpur: Winter 2018

Project mentor, WSAI Summer Internship Program, IIT Madras

Faculty guide: Prof. B. Ravindran.

• Madhangi Karimanal, B.E. (CSE), SSN College of Engineering (May to Jul 2026). Co-author of Pluralis v0.1.
• Paruchuri Sriram, M.Tech. (CSE), IIIT Hyderabad (May to Jul 2026). Project: Saakshya: A Multi-Agent, RAG-Enhanced Framework for Deepfake Detection and Governance.

Informal research mentoring

• Saumyaranjan Mohanty, Ph.D. student, IIT Hyderabad (advisor: Dr. Konda Reddy Mopuri). Joint papers: DiRe (WACV 2026) and X-AVDD (NeurIPS 2026).
• Mohd Wasif Raza, M.S. student (EE), IIT Madras. Co-first author on machine unlearning for federated socially fair k-means.
• Deeraj S K, B.Tech. student, SVNIT Surat, and WSAI summer intern (faculty guide: Dr. Saurav Prakash). Co-first author on the same paper.
• Akanksha Singh, Post-baccalaureate Fellow, WSAI, IIT Madras (faculty guide: Prof. B. Ravindran).
• Vraj Thakkar, Post-baccalaureate Fellow, WSAI, IIT Madras (faculty guide: Dr. Krishna Pillutla).

Invited research talks

• On MAP Inference of Ferromagnetic Potts Models and Nonsymmetric DPPs: IIT Madras, Department of Data Science and AI (Sep 2024), IIT Kanpur, Department of CSE (Jun 2024), IISc Bangalore (recording) (May 2024), IIT Hyderabad, Department of CSE (May 2024)
• On Clustering, MAP Inference, and Causal Inference: Broad Institute of MIT and Harvard (Nov 2022), Stanford University (Oct 2022)
• Beyond Perturbation Stability: University of Chicago Theory Lunch (Nov 2022)

Service

• Conference reviewer: AAAI (2026), AISTATS (2023 to 2024), ECCV (2026), FSTTCS (2024), ICLR (2022 to 2024), ICML (2022 to 2026), NeurIPS (2021 to 2025)
• Outreach: Workshop Leader and Mentor, NeurIPS 2022 High-School Outreach Program; invited talks for school students at RSI-C Chennai (May 2026) and Oak Valley International School, Hyderabad (Jun 2025)
• Community: Board Member, CS Ph.D. Advisory Council, Northwestern University (2020 to 2022); Webmaster, Northwestern CS Theory group (2018 to 2023)