Pratinav
Seth

Lead Research Scientist · Lexsi Labs · Model Science group, India & Paris (remote)

I build AI systems you can trust: how models are trained, how they work inside, and how to keep them safe once they are fine-tuned, compressed and put to work as agents.

40Peer-reviewed papers
290+Citations · h-index 9
9Libraries built or led
350+GitHub stars

Open to research collaborations and talks. If you work on post-training or interpretability, I’d like to hear from you.

About

I lead the Model Science group at Lexsi Labs, more than 30 people in India and Paris, and founded its AI Safety & Alignment team. We work on post-training, alignment, and mechanistic interpretability for language models and agents.

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. Nine libraries I designed or led have 350+ 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 40 peer-reviewed papers and 290+ 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.

  1. 2020–24Mars Rover ManipalTrainee to senior researcher leading the AI Research Wing; mentored 10+ undergraduates
  2. 2022–23IIT KharagpurMedical image analysis and explainable AI for chest radiographs, KLIV Lab
  3. 2023Bosch ResearchGenerative data augmentation for autonomous-driving perception
  4. 2024Mila, Rolnick LabGeospatial climate AI; Alberta Wells Dataset (first author, ICML 2025)
  5. 2024–nowLexsi LabsResearch Scientist to Lead Research Scientist, Model Science

Full experience →

61 papers · 40 peer-reviewed · 9 libraries

Trustworthy AI, from the data to the deployed agent

A model passes through many hands before anyone relies on it: someone curates its data, trains and aligns it, compresses it for cheaper hardware, fine-tunes it for a new domain, and wires it into an agent that acts on its own. Each step can change what the model does. I work across that whole lifecycle, on the research that explains where things go wrong and on the systems that let teams fix them and prove the fix held.

Alignment & post-training

Curate the data, train the model, and when fine-tuning or quantization erodes safety, repair it in the weights or at inference time.

All papers →

Interpretability

Open models up: from post-hoc explanations to the circuits behind refusal, and edits that change them.

All papers →

Agents & evaluation

Train tool-using agents, test that safety holds when they act, and keep one auditable record of every change.

All papers →

Foundation models & applied AI

Tabular foundation models, and AI for health, climate and finance where reliability decides whether a model gets used.

See the areas ↓

See the systems behind this →

Applied across domains

2025–26 · 8 papers · 8 peer-reviewed

Foundation models for structured data

The Orion tabular foundation models and TabTune, plus distillation, ensembling, and deployments in health and credit risk.

All 8 →
2022–25 · 3 papers · 3 peer-reviewed

Climate & earth observation

Satellite and remote-sensing AI: finding abandoned oil and gas wells, contrail detection, and multispectral marine-debris segmentation.

All 3 →
2022–25 · 4 papers · 4 peer-reviewed

Medical imaging

Glaucoma detection from retinal context, skin-tone bias in skin-lesion classifiers, brain-tumor segmentation and diabetic retinopathy.

All 4 →
2023–26 · 3 papers · 3 peer-reviewed

Vision & super-resolution

Guided thermal super-resolution with Laplacian and contrastive fusion networks, from a CVPR workshop to Nature Scientific Reports.

All 3 →
2023–25 · 5 papers · 5 peer-reviewed

NLP for social good

Detecting mental manipulation in conversations, LLMs for farmers’ queries, online-sexism detection, Bangla sentiment and biomedical lay summaries.

  • SELF-PERCEPT · ACL 2025 (Main) · NAACL-SRW 2025
  • AgriLLM · NLP4PI Workshop, EMNLP 2024 · KDD 2024 Undergraduate Consortium
All 5 →
2021–23 · 3 papers · 2 peer-reviewed

Uncertainty & robustness

Predictive uncertainty under distribution shift, test-time-augmented ensembles, and how synthetic training data affects models.

All 3 →

Selected work

Nine libraries came out of this work: SafeTune, AlignTune, AgentTune, CircuitKIT, CuratorKIT, AuditKIT, TabTune, DLBacktrace and xai_evals. See them with their roles and star counts →

Flagship papers

Impact

Citations per year, Google Scholar, as of 2026-11-15. *Year to date.
Libraries and GitHub stars
LibraryRole★
TabTuneDesigned & built v1, then mentored144
AlignTuneLed & built; design lead & mentor49
CuratorKITDesigned & mentored37
DLBacktrace v2Architected & built28
CircuitKITLed & built32
SafeTuneLed & built24
xai_evalsBuilt (first author)18
AuditKITTeam lead & contributor19
AgentTuneDesigned, led & mentored14

News

Earlier (19)

Quick answers

What does Pratinav Seth work on?
Trustworthy AI for language models and agents: post-training and alignment, mechanistic interpretability, and repairing safety failures in model weights. He leads the Model Science group at Lexsi Labs.
What has he built and led?
The Lexsi Labs alignment and safety stack: 9 libraries he designed, built or led, including AlignTune, SafeTune, CircuitKIT, CuratorKIT, AgentTune and TabTune, with 350+ GitHub stars between them.
Where has he published?
61 papers, 40 of them peer-reviewed, at venues including ICML, ACL, EMNLP, WWW, MIDL and Nature Scientific Reports, with 290+ citations on Google Scholar.
How can I contact him?
By email at seth.pratinav@gmail.com or on LinkedIn. He is open to research collaborations and talks.

Let’s talk.

If you work on post-training safety, interpretability, or the systems that put them into production, write to me. I’d like to hear what you’re building.

Open to: Research collaborations · Talks & panels

Students & early-career researchers: I can occasionally advise. Tell me who you are, what you’re working on, and the one question you want help with.

Roles at Lexsi Labs: apply through lexsi.ai rather than asking me for a referral.