I work on modelling biological intelligence: representation learning for medical imaging, and agentic systems that reason over multimodal clinical data.
I'm a Project Research Scientist at the Health Informatics and Imaging Lab at IIT Madras, working with Prof. Pradeeba Sridar on ICMR-funded research in medical image analysis. My work centres on self-supervised representation learning for ultrasound — designing pretraining objectives around the physics and anatomy of the data rather than applying generic SSL — and, more recently, on agentic systems that reason over multimodal clinical evidence.
I'm interested more broadly in modelling biological intelligence: what biological systems can teach us about building better learning machines, and what learned representations can tell us about biology. Before returning to research I worked on product analytics at Ather Energy and computer vision and reinforcement learning at Samsung Research. I hold a dual B.Tech + M.Tech in Engineering Design from IIT Madras.
Research
Peritumoral masked autoencoding for axillary lymph node metastasis prediction
Nodal status in breast cancer is decided by tissue that sits outside the tumour, but most imaging models are trained to attend to the lesion itself. PS-MAE uses a Sobel-guided masking scheme over the peritumoral ring to force the encoder to represent the margin and surrounding tissue during pretraining, rather than treating it as background.
Under review, IEEE JBHI
A foundation model for breast and multi-anatomy ultrasound
A JEPA-style self-supervised model trained across a curated multi-source ultrasound corpus, using masking informed by ultrasound image formation rather than uniform random patches. The aim is a general-purpose ultrasound encoder that transfers across anatomies and acquisition settings.
In preparation, Medical Image Analysis
Wavelet-domain adversarial training for cross-population segmentation
Breast ultrasound segmentation models degrade when moved between scanner populations. WAAU-Net applies adversarial alignment in the wavelet domain, targeting the frequency bands where acquisition differences concentrate, to improve robustness across populations.
Under review, IEEE JBHI
Agentic multimodal decision support for prostate cancer — MICCAI 2026 Challenge
An agentic pipeline that integrates histopathology, clinical, and molecular evidence to produce prostate cancer risk assessments, evaluated on both prediction accuracy and the fidelity of the reasoning trace it produces.
Ongoing
Publications
Under review
- PS-MAE: Peritumoral masked autoencoding for axillary lymph node metastasis prediction. IEEE Journal of Biomedical and Health Informatics, 2026.
- WAAU-Net: Wavelet-domain adversarial training for cross-population segmentation. IEEE Journal of Biomedical and Health Informatics, 2026.
In preparation
- OpenBUS: A foundation model for breast and multi-anatomy ultrasound. Medical Image Analysis.
Peer-reviewed
- ICECCME paper. ICECCME, 2026.
- Automated Flexible Needle Trajectory Planning for Keyhole Neurosurgery Using Reinforcement Learning. IROS, 2022.
- LiteAR: A Framework to Estimate Lighting for Mixed Reality Sessions for Enhanced Realism. Advances in Computer Graphics (CGI), 2022.
Service
- Reviewer: MICCAI, International Journal of Biomedical Imaging