AI researcher working on mechanistic interpretability, adversarial robustness and white-box AI control.
Research experience
MARS V, Meridian AI Research Summer · Fellow
Lindley Lentati track: White-Box Techniques for AI Control. Cambridge
Trained linear compliance probes on residual-stream activations across four instruction-tuned models from three labs. Ran a red-team/blue-team co-evolution loop against them across three attack families; the main result is negative: jailbreak success and probe evasion are mutually exclusive. The deployment lesson is the false positive rate: probes do find the jailbreak direction, but every probe keeps a false positive rate, so deployment needs a monitor cascade rather than one probe. Extending into trained attack policies behind a six-stage de-risking protocol. Multi-GPU SLURM cluster (8×B200), shared codebase with one collaborator. The harness is open source: Whetstone.
Redarc Labs · Co-founder and Research Lead
Delhi
An AI safety research lab, co-founded with Manan Wadhwa; our joint papers come from it. We started with interpretability for existential risk (four accepted workshop papers in 2026, one under review). The focus now is control and interpretability for non-model capability gains: what an agent can do because of its harness rather than its weights.
Run a six-person research fellowship. Cohort 01 is built around one thesis: expressed confidence in LLMs is separable from internal knowledge. Mentoring three first-author workshop projects; two papers under review at NeurIPS 2026 workshops. Wrote and teach Redarc Fellows, an eight-session mechanistic interpretability curriculum taught with PyTorch hooks and Hugging Face directly, no wrapper libraries.
Secure AI Futures Lab · Special Projects Fellow
Delhi hub
Teaching the Redarc Labs mechanistic interpretability bootcamp in person, and running reading groups on the weekends.
Lexsi Labs
Tabular Foundation Models Lab
Capability research on tabular foundation models: attributing model behavior back to synthetic pretraining data. Two papers at the ATTRIB workshop, NeurIPS 2026, and one at FMSD, ICML 2026. Listed here.
Apart Research · Research Fellow
Mentor: Jason Hoelscher-Obermaier · lead author, accepted at TAIS 2026, Oxford
Measured geometric divergence of ViTs and ResNets under seven adversarial attack algorithms using Hutchinson's stochastic Hessian trace estimator. Found a five-order-of-magnitude discrepancy between input-space and parameter-space curvature, showing that standard robustness metrics misclassify internal model instability.
Artificial Intelligence Institute of South Carolina · Research Intern
Mentor: Dr. Amitava Das
Co-author of AMBEDKAR (arXiv:2509.02133, cited by MIT Technology Review). Designed the fairness-aware speculative decoding algorithm, which reduces religion bias by 77% and caste bias by 23% at 6% latency overhead.
Publications
Full descriptions on the AI safety and AI capability pages. Papers with Manan Wadhwa come from Redarc Labs.
Peer-reviewed and accepted
- 2026 Known by Construction: Calibrating Model-State Instruments Against Models Whose State Was Installed, Hidden, or Never Present New in ML, NeurIPS 2026. With Manan Wadhwa (equal contribution)
- 2026 Dead Salmons Have a False-Positive Rate: A Calibrated Null-Model Test for Interpretability Interpretability as a Science, NeurIPS 2026. With Manan Wadhwa (equal contribution)
- 2026 Steered Affect Moves a Partner Agent’s Activations, and the Logistic Probe That Reads It Returns a Tie-Break, Not an Estimate Interpreting Agent Behavior, NeurIPS 2026. With Wadhwa (equal contribution), Ghosh, Kumar, Garg
- 2026 Toxin Circuits in ESM-2: Mechanistic Interpretability for Biosecurity Screening GenBio Workshop, ICML 2026. With Manan Wadhwa (equal contribution)
- 2026 When Does Synthetic Relational Data Teach Models to Use Relations? Tracing Predictive Structure from Pretraining Data to Model Behavior ATTRIB Workshop, NeurIPS 2026. First author, with Bouadi, Bouarour, Kulkarni, Tanna, Sankarapu
- 2026 From Behavior to Provenance: Attributing Tabular Foundation Models to Synthetic Pretraining Data ATTRIB Workshop, NeurIPS 2026. With Bouadi, Bouarour, Tanna, Sankarapu
- 2026 What You Pretrain On Matters: Synthetic Task Distributions Determine Tabular Foundation Model Quality FMSD Workshop, ICML 2026, accepted poster. Bouadi, Bouarour, Kulkarni, Dubey, Tanna, Sankarapu; arXiv:2605.18971
- 2026 Loss Landscape Response to Adversarial Perturbation Is Architecture-Dependent TAIS 2026, Oxford. Lead author, with Jason Hoelscher-Obermaier
- 2026 Physics-Informed Neural Operators with Multi-Method Explainability for Ocean Temperature Prediction XAI4Science Workshop, AAAI 2026. Sole author
- 2025 Spectral Regularization as a Safety-Critical Inductive Bias Reliable ML Workshop, NeurIPS 2025. Sole author
- 2025 AMBEDKAR: A Multi-level Bias Elimination through a Decoding Approach with Knowledge Augmentation for Robust Constitutional Alignment of Language Models arXiv:2509.02133. Mukhopadhyay, Kasat, Dubey, et al. Cited by MIT Technology Review
Under review and in preparation
- 2026 Distance to the Boundary Is Not Steering Resistance Under review, UniReps. With Manan Wadhwa (equal contribution)
- 2026 Doubt Is a Decision: A Sparse Circuit Commits Language Models to “I Don’t Know” arXiv preprint, coming soon. With Mehta, Chadha, Singhal (equal contribution), Wadhwa
- 2026 Isolating Refusal and Compliance Heads: Unconditional Bipolar Defenses Safety circuits, Qwen2.5-7B
- 2026 Pre-commitment Probing in Open-Weight Reasoning Models Mechanistic interpretability, OLMo-3-7B-Think
- 2026 Context Poisoning in Tabular ICL: Geometric Hijacking and Inverse Scaling Mechanistic interpretability, tabular foundation models
Open-source tooling
- Harness whetstone Red×Blue harness for adversarial safety-monitor research. Built for the MARS fellowship.
- Package neuron_autolabel Auto-interpretability for LLaMA-style models, scored by detection AUC rather than simulation.
- Method dead-salmon-fpr Calibrated null-model test for interpretability claims.
- Tool mechanistic-tamper-verification Two-tier integrity verifier for tampered model updates.
Education
BS in Data Science
Indian Institute of Technology Madras
Focus: AI safety, interpretability, fairness, governance.
B.Tech in Mechanical and Automation Engineering
Maharaja Agrasen Institute of Technology, Delhi
Teaching
Redarc Fellows
Eight-session mechanistic interpretability cohort
Written and taught by me, with PyTorch hooks and Hugging Face directly and no interpretability library in the stack. The curriculum is here.
Technical skills
Python, CUDA, MATLAB, Bash, JavaScript
PyTorch, JAX, TransformerLens, NNsight, Hugging Face, DevInterp, LangChain, Docker, SLURM (multi-node, 8×B200)
QK/OV circuit decomposition, direct logit attribution, activation patching, steering vectors (CAA), logit lens, linear probing, SAE feature attribution, nanoGCG and suffix-gradient attacks, Hutchinson Hessian trace, SVD and spectral dynamics