About & Research Interests
I’m a final-year B.Tech student at the University of Mumbai and currently a research intern at A*STAR Singapore. While I've always been fascinated by what artificial intelligence can achieve, my primary focus has shifted to a much more urgent question: how can we actually trust what these models are doing?
Why AI Safety Matters to Me: We are watching AI capabilities accelerate at an incredible pace, often exceeding human performance in complex domains. But as these models become more capable, they also become more opaque. It deeply concerns me that we routinely deploy "black box" systems into high-stakes environments—like healthcare—without truly understanding how they make decisions. High accuracy on a benchmark doesn't mean a model is reasoning correctly; often, it’s just finding clever, superficial shortcuts. I believe the growing gap between AI's rapid advancement and our ability to safely interpret it is one of the most critical challenges of our time.
My Approach: I want to move our field beyond just chasing higher accuracy scores. My research sits at the intersection of Technical AI Safety and Mechanistic Interpretability. Instead of taking a model's output at face value, I try to open the black box. I build rigorous algorithmic audits—using matched counterfactual probes, causal interventions, and Sparse Autoencoders (SAEs)—to untangle the actual reasoning pathways inside the network. My goal is to catch dangerous failure modes and make them legible to developers before a model is deployed.
What I'm Working On: Right now, I’m investigating multimodal foundation models (LVLMs). In my recent work, I found that even when a medical AI achieves exceptional scores, it often isn't actually reading the clinical text; it's just relying on pre-training biases and lexical scaffolding. By isolating these exact shortcut mechanisms at the activation layer, I hope to provide the auditing tools necessary to catch unsafe models and ultimately guide the development of AI we can genuinely rely on.
Publications
Reading Between the Lesions: Auditing Whether Multimodal Dermatology Classifiers Actually Use Clinical Text
A mechanistic audit asking a safety-critical question: when a multimodal classifier scores well, is clinical text actually driving the decision—or are priors, token scaffolding, and shortcuts? I designed matched counterfactual probes, ablations, and Top-K sparse autoencoders on fused representations to investigate this.
- Semantic grounding fails: Aligned clinical text decreased the correct-class margin (−0.0102 mean shift).
- Text presence ≠ text meaning: Ablation probes expose reliance on lexical scaffolding over semantics.
- Foundation-model collapse: LLaVA-Med shows 0% prediction flips under explicitly contradictory histories.
Causal Auditing of Latent Affect in Language Models via Activation Steering
Mechanistic interpretability and inference-time steering without retraining. This project involves applying mechanistic interpretability to isolate internal failure modes and causal drivers within the model's hidden layers, and developing a steering mechanism to enforce model safety constraints.
View PreprintBeyond Accuracy: An Interpretability-Driven Audit of Deep Learning Models for Pneumonia Detection from Chest X-Rays
Artifact-driven shortcut discovery in clinical imaging.
View ECR 2026 PosterManifold-Constrained Counterfactual Explanations for Medical Diagnostics
Biologically constrained counterfactuals for diagnostic verification.