About

I started in computer vision and medical imaging — uncertainty quantification, model fairness, and segmentation with teams at IIT Kharagpur and Bosch Research. Moving into language models, I kept running into the same problem: a model that looked aligned on benchmarks would fail in ways that weren’t visible from the outside. That gap — between what training instils and what survives deployment — became the question I keep returning to.

On the safety side, I study how fine-tuning and quantization silently erode alignment, and how to find and repair the specific weights responsible using mechanistic interpretability: circuit-level refusal (C-ΔΘ), quantization-permanent unlearning (Forgetting That Sticks), and safety drift auditing. On the structured data side, I work on tabular foundation models — how to train, fine-tune, and distil them down to something deployable (TabTune, Orion-MSP, Orion-BiX). Most of this ships as open-source tools. The next problems I care most about are safety in reasoning models and interpretability for agentic systems.

I lead the Model Science group at Lexsi Labs across India and Paris. I completed my B.Tech in Data Science from MIT Manipal in 2024, and before that worked at Mila Quebec AI Institute (with Prof. David Rolnick), Bosch Research India, and IIT Kharagpur. I’m an AAAI Undergraduate Consortium Scholar (2023 → Mentor, 2026).

Recent work: AlignTune · CircuitKIT · C-ΔΘ · CuratorKIT · SafeTune · Forgetting That Sticks · TabTune · Orion-MSP

Feel free to reach out or see my Resume.

Key Highlights

Published at ICML, ACL, WWW, MIDL, IJCNN, AAAI, and Nature Scientific Reports, plus workshops and shared tasks at CVPR, NeurIPS, ICLR, MICCAI, EurIPS, ACL, and SIGMOD.

• Research Focus: LLM safety post-training — circuit-level mechanistic interpretability, safety weight arithmetic, unlearning, post-training alignment, and a growing focus on AI agents & evaluation.
• 7 open-source libraries released — see Libraries & Toolkits.
• AAAI Undergraduate Consortium Scholar & Mentor (2023 → 2026); Spotlight Talk at EurIPS Workshop on Private AI Governance, Copenhagen (December 2025).

News

Recent Publications & Acceptances

Academic Service & Reviewing