I am a computational postdoctoral fellow in the Koo Lab at Cold Spring Harbor Laboratory. My research makes AI systems transparent, from attribution methods in computer vision to generative models that design DNA with tunable regulatory activity.
01 — About
I aim to bridge human and machine intelligence, creating AI that is robust in the wild and transparent enough to be a trusted partner in scientific discovery.
At Cold Spring Harbor Laboratory, I am part of the Koo Lab within the Simons Center for Quantitative Biology, where my research with Dr. Peter Koo centers on generative modeling for genomics: designing regulatory DNA with tunable activity (D3) and developing inference-time, model-agnostic methods that push sequence generation beyond the activities seen during training (GPA).
Before CSHL, I was a postdoctoral associate in the Sinha Lab for Developmental Research within MIT's Brain and Cognitive Sciences department, where Dr. Xavier Boix, Dr. Pawan Sinha, and I studied explainability under distribution shift.
My Ph.D., in Computer Science & Engineering at IIT Hyderabad under Dr. Vineeth N Balasubramanian, focused on interpretability and robustness for computer vision: self-explaining networks, and defenses against unseen, adversarial, and attributional attacks.
02 — Research
One thread runs through my work: first understand what a model knows, then turn that understanding into action, to design, to steer, to reach past the limits of the training distribution (GPA), and increasingly to imagine data that does not exist yet.
Attribution and self-explaining models that reveal what a vision network sees and why: Grad-CAM++, ante-hoc concept models, attributional robustness, and more.
Understanding and debugging vision models when the test world drifts away from the world they were trained on.
Turning understanding into design: generating regulatory DNA with tunable activity through discrete diffusion (D3).
Pushing generation beyond the activities seen in training, at inference time, with no retraining and no model-specific hooks, so it drops onto frozen generators (GPA).
An open problem I'm drawn to: aligning different medical modalities, histopathology, regulatory DNA, transcriptomes, around a shared, interpretable space of concepts. Once they share that language, one modality can be generated from another, concept for concept, then interrogated with counterfactuals: what changes if a single concept does?
It ties together the threads I care about most, interpretability, generative modeling, and causal intervention, with real stakes for biology and medicine.
// early and unproven, but the kind of problem worth the next few years.
03 — Highlights
Recognition
Updates
“GPA: Generative Population Annealing for Test-Time Sequence Design” accepted as a poster at GenBio 2026, the ICML Workshop on Generative & Agentic AI for Biology.
Presented “Understanding DNA Discrete Diffusion for Engineering Regulatory DNA Sequences” at the Workshop on AI for Nucleic Acids, ICLR 2025.
Long oral “Designing DNA With Tunable Regulatory Activity Using Discrete Diffusion” at MLCB 2024; also featured at the NeurIPS 2024 Workshop on AI for New Drug Modalities.
Started as computational postdoctoral fellow at Cold Spring Harbor Laboratory.
Presented ongoing work on explainability under distribution shift at Fujitsu Limited, Japan.
Started as postdoctoral associate at MIT.
Successfully defended my Ph.D. dissertation.
04 — Trajectory
Experience
Generative modeling for genomic sequence design.
Explainability under distribution shift.
Self-explaining neural networks with meaningful concepts as building blocks.
Causal inference and applications of causality in machine learning.
Education
Rational Deep Machines: toward explainable, trustworthy, and robust deep learning systems.
Source camera identification: classifier learning, learning curves, and their interpretation.
05 — Publications
GPA: Generative Population Annealing for Test-Time Sequence Design
Leveraging Test-Time Consensus Prediction for Robustness against Unseen Noise
Source Camera Identification Model: Classifier Learning, Role of Learning Curves and their Interpretation
06 — Talks & Workshop Papers
07 — Contact