Pratinav
Seth

I lead the Model Science group at Lexsi Labs: more than 20 researchers and interns in India and Paris working on post-training, alignment, and mechanistic interpretability for language models.

My research asks what survives post-training. Fine-tuning, quantization, and deployment all change how a model behaves, and they can erode its safety. I look for the mechanism behind each failure and repair it in the weights: circuit-restricted refusal edits (C-ΔΘ), safety-drift auditing and repair (SafeTune), and alignment transfer at inference time (ALIGNBEAM).

I also build the systems that make this repeatable. I architected Lexsi's alignment and safety stack and scaled AlignTune into the multi-backend post-training system behind our BFSI, legal, and healthcare work. Seven open-source libraries I designed or helped build have 300+ GitHub stars between them.

Before Lexsi, I worked with David Rolnick at Mila on climate AI, where my bachelor's thesis became the Alberta Wells Dataset (first author, ICML 2025). Earlier, I built generative data augmentation for autonomous-driving perception at Bosch Research and rose to lead the AI Research Wing at Mars Rover Manipal.

I have 41 peer-reviewed papers and 220+ citations, at venues including ICML, ACL, EMNLP, WWW, MIDL, and Nature Scientific Reports. I was an AAAI Undergraduate Consortium Scholar in 2023 and came back as a mentor in 2026.

Pratinav Seth is a Lead Research Scientist at Lexsi Labs, where he leads the Model Science group across India and Paris. His research studies how fine-tuning, quantization, and deployment change the behavior of language models, and how to find and repair the circuits behind safety failures. He architected Lexsi's alignment and safety stack, including AlignTune, CircuitKIT, and SafeTune. He has 41 peer-reviewed papers and 220+ citations, at venues including ICML, ACL, EMNLP, WWW, and Nature Scientific Reports. Before Lexsi, he worked with David Rolnick at Mila Quebec AI Institute and at Bosch Research. He holds a B.Tech in Data Science & Engineering from Manipal Institute of Technology and was an AAAI Undergraduate Consortium Scholar (2023) and Mentor (2026).

I lead the Model Science group at Lexsi Labs, directing research on post-training, alignment, and mechanistic interpretability. The group has grown to more than 20 researchers and interns across India and Paris, and alongside the research I handle hiring, grants, the roadmap, and getting results into the product.

The research. Fine-tuning, quantization, and deployment all change how a model behaves. I study which of those changes break safety, find the mechanism responsible, and test targeted ways to fix it. That thread runs through circuit-restricted refusal edits (C-ΔΘ), circuit discovery and attribution (CircuitKIT), safety-drift auditing and repair (SafeTune), unlearning that survives quantization (Forgetting That Sticks), and alignment transfer at inference time (ALIGNBEAM). I'm now extending it to agents and to evaluation.

The systems. I architected Lexsi's alignment and safety stack: single-purpose libraries for data curation, post-training, interpretability, repair, compression, and evaluation that all write to one shared evaluation record. I scaled AlignTune into the multi-backend post-training system behind our BFSI, legal, and healthcare work. Before the LLM work, I led early design and benchmarking for the Orion tabular foundation models and architected TabTune, with follow-on work on distillation, ensembling, health data, and credit risk.

Before Lexsi. At Mila Quebec AI Institute I worked with David Rolnick on climate AI; my bachelor's thesis became the Alberta Wells Dataset, a satellite benchmark for finding abandoned oil and gas wells (first author, ICML 2025). At Bosch Research I built a latent-diffusion pipeline that generated hard training examples for autonomous-driving perception. At IIT Kharagpur's KLIV Lab I worked on medical imaging and explainable AI for chest X-rays.

As an undergraduate. I did research on healthcare NLP at CUREYA, synthetic tabular data at Eedge.ai, malware classification at MAHE, metric learning with NEC Labs and IIT Roorkee, and agricultural NLP with Wells Fargo. I rose from trainee to lead the AI Research Wing at Mars Rover Manipal, mentoring more than ten undergraduates; co-founded The Data Alchemists; and co-led Research Society MIT Manipal, a 90-member research group.

In numbers. 41 peer-reviewed papers and 220+ citations, at ICML, ACL, EMNLP, WWW, MIDL, IJCNN, and Nature Scientific Reports, plus workshops at NeurIPS, ICLR, CVPR, and MICCAI. Seven open-source libraries with 300+ GitHub stars. I review for AAAI, NeurIPS, CVPR, and ECCV, and I was an AAAI Undergraduate Consortium Scholar in 2023 and a mentor in 2026. B.Tech in Data Science & Engineering, Manipal Institute of Technology, 2024.