Pratinav Seth

Experience

I design and ship systems that keep LLMs and agents safe and aligned through fine-tuning and deployment, and I lead the research behind them.

5+Years doing research
20+Researchers & interns led
9Libraries built or led
19Papers as lead author
34Papers as senior author

Systems I’ve designed

The Lexsi alignment & safety stack. I architected and lead a set of composable libraries that keep enterprise LLMs and agents safe and aligned through fine-tuning, compression and deployment. Each library does one job, and all of them write to one shared, reproducible evaluation record. Nothing ships until that record shows the fix kept the model's useful behavior.

  1. Curate

    Data curation and synthetic generation for post-training, with provenance for every example

  2. Post-train

    Modular SFT, preference and RL training for models and tool-using agents, behind one interface

  3. Inspect

    Find where a behavior lives: circuit discovery, attribution, interchange interventions, post-hoc explanation

  4. Repair

    Audit safety drift and correct it with small, verified, circuit-restricted weight edits

  5. Compress

    Structured pruning and quantization that keep safety and capability intact

    • Pruning libraryBuilt (internal)
    • Quantization libraryDesigned & mentored (internal)
  6. Evaluate

    Safety, utility and side-effects on one reproducible record, with LLM judges and cost tracking

Audit, correct in the weights, re-verify. We check every repair for capability loss and side effects before release.

The libraries

Designed & built v1, then mentored★ 137

TabTune

Unified inference, fine-tuning and benchmarking for tabular foundation models (21 models in 7 architecture families).

GitHubPaper

Led & built; design lead & mentor★ 44

AlignTune

Modular post-training toolkit for SFT, DPO, GRPO and RLHF with interchangeable backends; runs production alignment work in BFSI, legal and healthcare.

GitHubPaper

Designed & mentored★ 30

CuratorKIT

Provenance-grounded data curation and synthetic generation for LLM post-training.

GitHubPaper

Architected & built★ 28

DLBacktrace v2

torch.export-based, model-agnostic explainability for LLMs and MoEs with CUDA acceleration.

GitHubPaper

Led & built★ 26

CircuitKIT

Circuit discovery, evaluation and application toolkit for mechanistic interpretability.

GitHubPaper

Led & built★ 19

SafeTune

Audits and repairs safety drift in fine-tuned LLMs.

GitHubPaper

Built (first author)★ 18

xai_evals

Framework for evaluating post-hoc local explanation methods.

GitHubPaper

Contributor★ 12

AuditKIT

Source-available evaluation library for benchmarks, LLM-as-judge, RAG, model comparison, cost metrics and experiment tracking.

GitHubWebsite

Designed, led & mentored★ 10

AgentTune

Source-available training lifecycle for building, training, evaluating and maintaining tool-using LLM agents.

GitHubPaper

How it’s built

  • Single-responsibility libraries that compose through a shared evaluation record, instead of one monolith.
  • Internal/public twins: each library has an internal build for client work and a public release, plus a lightweight studio app for demos.
  • Config-driven, reproducible evaluation: judge models, tiny-model CPU tests and CI on every library, verified across CUDA and Apple MPS.
  • Research → benchmark → feature: a research question becomes a benchmark, and the benchmark becomes a library feature.

Core stack

PyTorchHugging Face TransformersPEFT / LoRASFT, DPO, GRPO, RLHFvLLMCircuit discovery & attributionActivation steering & weight editingSafety evaluation & red-teamingBenchmark designQuantization (AWQ, bitsandbytes)Structured pruningState-space / hybrid modelsTabular foundation modelsAgent & tool-use orchestrationCUDA / MPSCI & release engineering

Research positions

2026.04 – Present

Research and engineering lead for AI safety & alignment

Lead Research Scientist at Lexsi Labs, Lexsi.ai (Aurionpro Solutions Group), Remote

  • Platform architecture: Architected the lab's alignment & safety stack: 9 single-purpose libraries for data curation, post-training, interpretability, safety repair, compression and evaluation, all writing to one shared evaluation record. The open-source ones have 300+ GitHub stars.
  • Post-training platform: Led and built AlignTune, then scaled it as design lead into the multi-backend post-training system (SFT, DPO, GRPO, RLHF-style training, model merging) behind the lab's BFSI, legal and healthcare work.
  • Safety repair: Led and built SafeTune (EMNLP 2026 Demo), which audits safety drift after fine-tuning and repairs it with small weight edits checked for capability loss. Co-led C-ΔΘ, circuit-restricted weight edits for selective refusal (ICML 2026 workshop).
  • Interpretability & agents: Led and built CircuitKIT for circuit discovery, attribution and interventions. Designed and led AgentTune for training tool-using agents, mentoring the team that built it.
  • Compression & evaluation: Built the lab's pruning library and designed its quantization library, with checks that compressed models keep their safety behavior. Set up config-driven evaluation with LLM judges and CI on CUDA and Apple MPS; contributed to AuditKIT.
  • Research output: Of the group's 11 papers accepted at EMNLP 2026 and NeurIPS 2026 workshops, I was first author on 4 and mentored the first authors of the other 7, including ALIGNBEAM (EMNLP Main) and Provenance-Grounded Gating (EMNLP Findings). Applied the work to regulatory-compliance, financial-services, insurance and Indic-language deployments.
  • Leadership: Lead the Model Science group, 20+ researchers and interns across India and Paris: research direction, hiring, grants and roadmap. Set the team's standards for evidence-first measurement and reproducible evaluation, and review code across engineering, including the lab's internal AI coding-agent platform.
14 papers from this role
2025.07 – 2026.03

Post-training, interpretability and tabular foundation models

Research Scientist at Lexsi Labs, Lexsi.ai (Aurionpro Solutions Group) [Previously: AryaXAI, Arya.ai], Remote

  • Team: Grew the Model Science group into the lab's core team for alignment, safety and interpretability: 14 interns and 5 full-time researchers across India and Paris. Ran grants, hiring and planning.
  • Post-training: Led and built AlignTune, a modular post-training toolkit covering SFT, preference optimization and safety methods. Senior author on Interpretability as Alignment (EurIPS 2025 workshop, spotlight talk), which set the team's design principle.
  • Interpretability: Architected and built DLBacktrace v2 on torch.export graph capture, so it works across model architectures, with CUDA acceleration for LLMs and MoEs. Applied interpretability to compression with interpretability-aware pruning for medical imaging (MICCAI 2025 workshop).
  • Tabular foundation models: Led problem framing, early design and benchmarking for the Orion models (Orion-BiX at WWW 2026, Orion-MSP at an EurIPS 2025 workshop) and their extension to regression.
  • TabTune: Designed and built the first version of TabTune (WWW 2026 Demo) with a companion fine-tuning study, then mentored follow-on work on distillation and ensembling (ICML 2026 workshops) and credit-risk prediction (SIGMOD 2026 workshop, oral).
  • Compression: Built the lab's pruning toolkit and brought interpretability-guided compression into the production deployment pipeline.
  • Talks: Gave a spotlight talk at EurIPS 2025 and presented a poster at a MICCAI 2025 workshop.
6 papers from this role
2024.07 – 2025.06

Explainability methods and XAI evaluation

Research Scientist at AryaXAI Alignment Labs (rebranded to Lexsi Labs in 2025), Remote / Mumbai, India

  • Focus: Interpreted black-box models and measured how reliable XAI methods are for high-stakes uses such as fraud detection; started the lab's work on tabular foundation models.
  • Explainability: Generalized DL-Backtrace to work on any model architecture and mentored its IJCNN 2025 paper. Built xai_evals (first author), a framework for benchmarking post-hoc explanation methods.
  • Interpretability-guided optimization: Explored post-hoc optimization and alignment methods guided by interpretability, across model architectures.
  • Leadership: Mentored two research interns, led hiring of interns and full-time scientists in Paris and India, wrote technical documentation for stakeholders, and started proof-of-concept projects for the lab's methods.
  • Representation: Represented R&D in client engagements and presented AryaXAI's work at industry forums, including the 5th MLOps Conference.
3 papers from this role
2024.01 – 2024.06

Climate AI: finding oil and gas wells from satellites (ICML 2025)

Research Intern at Rolnick Lab, Mila Quebec AI Institute, Remote

  • Project: Computer vision and deep learning for geospatial applications targeting climate change
  • Focus: Detecting abandoned oil and gas wells from satellite imagery; created new geospatial dataset and benchmarked deep learning models
  • Mentor: Dr. David Rolnick (McGill University, Université de Montréal, Mila)
  • Outcome: Led to ICML 2025 publication on Alberta Wells Dataset
2023.06 – 2023.10

Generative data augmentation for autonomous driving

Computer Vision Research Intern at Robert Bosch Research and Technology Center India, Bangalore

  • Project: Vision-based generative AI for autonomous driving using Latent Diffusion Models
  • Focus: Generating additional data for difficult or misclassified samples to improve downstream task network optimization
  • Mentors: Mr. Koustav Mullick (CR/RDT-2), Dr. Amit Kale
2021.06 – 2023.12

Led an undergraduate AI research wing

Researcher, AI Research Wing at Mars Rover Manipal, Manipal, India

  • Progression: Rose from Student Trainee to Senior Researcher leading the AI Research Wing, driving the team's technical roadmap across computer vision, multimodal AI, and medical imaging
  • Recruitment & Training: Spearheaded recruitment that raised applicant quality; designed and ran training programs for incoming ML researchers
  • Mentorship: Mentored 10+ undergraduates (12+ total collaborators) under Dr. Ujjwal Verma; co-authored their publications, including workshop papers at NeurIPS, CVPR, ICLR, and MICCAI, an AAAI student abstract, and a Nature Scientific Reports journal paper
2023.04 – 2023.12

Fairness in skin-lesion classification

Undergraduate Researcher under Dr. Abhilash K. Pai, Dept. of Data Science & Computer Application, MIT MAHE, Manipal, India

  • Medical AI & Fairness: Studied the impact of pretraining techniques on skin-tone bias in skin lesion classification, funded by the MAHE Undergraduate Research Grant
  • Outcome: Led to a publication at the Pre-Train Workshop, WACV 2024

Research collaborations

2024 – 2026

Medical AI and LLM safety collaborations

Independent Research Collaborator with academic collaborators across IIT Kharagpur, IISc Bangalore, and MAHE, Remote

  • Medical AI: Lesion-aware MAE for glaucoma detection from retinal context, with Siddhant Bharadwaj and Chandra Sekhar Seelamantula (MIDL 2025 Short Papers)
  • LLM Safety: Introspective LLM detection of multi-person mental manipulation in conversations, with Danush Khanna, Sidhaarth Sredharan Murali, and collaborators (ACL 2025 & NAACL-SRW Workshop)
2023.12 – 2024.01

LLMs for farmers’ questions (AgriLLM)

Research Collaborator with Dr. Amit Agarwal, Wells Fargo AI Center of Excellence, Remote

  • AgriLLM: Built seq2seq LLMs (BART, T5, Flan-T5) for agricultural queries from Indian farmers on a highly noisy real-world dataset
  • Outcome: Published at the NLP4PI Workshop, EMNLP 2024, and the Undergraduate Consortium at KDD 2024
2022.05 – 2023.12

Chest X-ray screening and interpretability

Research Intern at KLIV Research Group, IIT Kharagpur (PI: Dr. Debdoot Sheet; mentored by Rakshith Satish), Remote

  • Healthcare AI: Developed an attention-driven dynamic Graph Convolutional Network for noisy, multi-label, comorbidity-aware chest radiograph screening
  • Interpretability: Investigated sanity checks for class activation maps in multi-label chest X-ray classification
2022.10 – 2023.03

Metric learning research

Research Collaborator with Dr. Vijay Kumar BR, NEC Labs & IIT Roorkee, Remote

  • Focus: Deep metric learning with self-supervised and contrastive learning; vision-based attention models and hyperbolic representation learning

Leadership

2022.08 – 2023.09

Ran a 90-member undergraduate research society

Co-President & AI Research Mentor, Research Society MIT Manipal, Manipal, India

  • Organization: Led a 90+ member undergraduate research organization across 10+ technical domains; managed recruitment from 250+ applicants and ran member-development pathways
  • Mentorship: Mentored 10+ undergraduates and co-authored the RSM-NLP team's shared-task papers
2022.11 – 2023.09

Co-founded an AI/ML club

Co-founder & Head of Artificial Intelligence and Machine Learning, The Data Alchemists, Manipal, India

  • Co-founding: Founded and led the AI/ML club, growing it to 30+ members and establishing a lasting community of practice
  • Programming: Ran workshops and technical events on ML fundamentals and projects

Early career

2022.06 – 2022.09

Malware classification (5th of 134 teams)

Undergraduate Research Assistant under Dr. Vidya Rao & Dr. Poornima P.K., Dept. of Data Science & Computer Application, MIT MAHE, Manipal, India

  • Cybersecurity & AI: Multi-class malware classification at the intersection of cybersecurity and AI
  • Outcome: Placed 5th of 134 teams in the 13th International Cyber Security Data Mining Competition
2022.03 – 2022.05

Synthetic tabular data

Machine Learning Intern at Eedge.ai, a deep-learning talent-design platform, Remote

  • Synthetic Data & Interpretability: Built generative models for synthetic tabular data to fine-tune downstream models; analyzed the effect of synthetic data on the interpretability of the fine-tuned models
  • Mentor: Ananta Mahapatra, CTO & Co-Founder
2022.01 – 2022.02

Healthcare NLP

Data Science (NLP) Intern at CUREYA, Aspexx Health Solutions, Remote

  • Healthcare AI: Built NLP pipelines for Reyana, a conversational healthcare chatbot: text classification, summarization, BERT models, and LSTM-based NLG for clinical dialogue