I lead the Model Science group at Lexsi Labs, directing research across post-training, alignment, mechanistic interpretability, and tabular foundation models. The team has grown to 20+ researchers and interns across India and Paris; I’ve worked on hiring, roadmap, grants, and research-to-product integration alongside the research itself.

I spend most of that time asking what survives post-training: how fine-tuning, quantization, and deployment change a model’s behavior, and how to find the specific mechanisms responsible. I use mechanistic interpretability to answer this, then audit, repair, or steer those mechanisms directly: circuit-level refusal (C-ΔΘ), circuit discovery and attribution (CircuitKIT), and safety-drift auditing and repair (SafeTune). Alongside this, I build the post-training infrastructure that ships it: AlignTune for training, CuratorKIT for data curation, and ALIGNBEAM for inference-time alignment transfer. I used to work on the Orion tabular foundation model series and TabTune too, with follow-on research in distillation, ensembling, and applications in structured health data and credit risk. I’m increasingly extending this work to agentic systems and evaluation.

Before Lexsi, I worked with David Rolnick at Mila Quebec AI Institute on geospatial climate AI. My bachelor’s thesis became the Alberta Wells Dataset, a satellite-imagery benchmark for detecting abandoned oil and gas wells, published first-author at ICML 2025. Before that, at Bosch Corporate Research, I built a generative data-augmentation pipeline for safety-critical autonomous-driving perception, using latent diffusion models to synthesize hard-negative training samples. Earlier, at IIT Kharagpur’s KLIV Lab, I worked on medical image analysis and explainable AI for chest radiographs.

Before all of that, I was already doing research as an undergraduate. I worked on NLP for a healthcare chatbot at CUREYA, synthetic tabular data at Eedge.ai, cybersecurity malware classification at MAHE, deep metric learning with NEC Labs and IIT Roorkee, and agricultural NLP with Wells Fargo. I also mentored 10+ undergraduates as a researcher at Mars Rover Manipal, rising from student trainee to lead its AI Research Wing; co-founded The Data Alchemists, growing it to 30+ members; and co-led Research Society MIT Manipal, a 90-member undergraduate research organization.

I’ve published 48 papers (31 peer-reviewed, 210+ citations) at ICML, ACL, WWW, MIDL, and Nature Scientific Reports, alongside workshops and shared tasks at NeurIPS, ICLR, CVPR, MICCAI, EurIPS, EMNLP, WACV, and ACM SIGMOD. I’ve released seven open-source libraries (230+ combined GitHub stars): TabTune, AlignTune, DLBacktrace v2, xai_evals, CircuitKIT, CuratorKIT, and SafeTune. I review for AAAI, NeurIPS, CVPR, ECCV, and other major venues, and I’m an AAAI Undergraduate Consortium Scholar (2023) turned Mentor (2026). I completed my B.Tech in Data Science & Engineering at Manipal Institute of Technology in 2024.