I build generative and foundational models for medical AI, and study how neurosymbolic and Bayesian structure can make machine reasoning more faithful. Currently a doctoral researcher at the Geisel School of Medicine, Dartmouth College.
I'm a PhD student in Quantitative Biomedical Sciences at the Geisel School of Medicine, Dartmouth College. My work sits at the meeting point of generative modeling, medical imaging, and structured machine reasoning — using deep generative models to make better clinical predictions, and probabilistic and symbolic structure to make those models more trustworthy.
For my dissertation I work with Indrani Bhattacharya on generative image models and foundational models for medical AI — including synthesizing PET imagery from CT to improve cancer-lesion segmentation. Across my doctoral rotations I built a CT organ-segmentation and metastasis-classification framework with Saeed Hassanpour, and developed PRISM-BN with Soroush Vosoughi, a collaboration I continue today.
Before Dartmouth I worked across industry and academic research — an AI founding team in Bangalore, the Indian Institute of Science, Osaka University, and Arizona State University's Banerjee Lab with the Mayo Clinic. My background runs from vision-language models and NLP to statistical methods for longitudinal clinical data. More on Google Scholar.
Deep generative models — pix2pix and beyond — for cross-modality synthesis (CT → PET) and downstream lesion segmentation. The dissertation core: generative and foundational models built for clinical imaging tasks.
Eliciting parameterized Bayesian networks from text, and probing how symbolic structure changes the reasoning of vision-language and language models. Where statistical structure meets learned representation.
Organ segmentation with TotalSegmentator feeding CT classification by metastasis origin and diagnosis category — comparing segmented vs. full-volume input and single- vs. multi-task heads, with interpretability throughout.
Hierarchical random-effects extensions of time-lag models for intensive longitudinal clinical data — recovering person-specific dynamics that group-level methods average away.
A corpus of 5,054 paired (text, parameterized Bayesian network) examples, built through a marginal-first CPD elicitation pipeline that extracts nodes, states, edges, and conditional distributions across multiple domains. Benchmarks six LLMs against a fixed judge; a central finding is that structural and CPD-level fidelity are decoupled.
An image-translation pipeline synthesizing PET imagery from CT to strengthen cancer-lesion segmentation, operating on preprocessed volumetric arrays rather than rendered images. Achieved SSIM 0.88 and FID 0.0145, with synthetic PET now being integrated into the segmentation workflow.
A dual-head ResNet framework classifying CT scans by the origin of skeletal metastases and diagnosis category, comparing segmented vs. full-volume input and single- vs. multi-task learning, with Grad-CAM interpretability and NIfTI export.
Scene-graph–augmented compositional reasoning for vision-language models, evaluated on Winoground. The work reframes a counter-intuitive result — that a strong base VLM can outperform scene-graph-injected variants — as a capability–augmentation tradeoff.
A hierarchical random-effects extension of the Differential Time-Varying Effect Model, giving each person their own lag structure via a Bayesian state-space VAR (Stan) and a continuous-time Ornstein–Uhlenbeck model (ctsem). Directly implements the random-effects extension named as future work by the original DTVEM authors.
Proceedings of WACV 2025 · pp. 2808–2817 peer-reviewed
European Conference on Technology Enhanced Learning (ECTEL) · Springer · pp. 504–517 best-paper nominee
arXiv:2306.14875
Physiological Measurement · 43(12): 124004 · IF 3.2 with Mayo Clinic
Graduate Research Assistant · Geisel School of Medicine
Geisel School of Medicine, Dartmouth College
Research Scientist · AI founding team
Research Fellow · with Prof. Soma Biswas
Visiting Researcher · with Prof. Hajime Nagahara
Research Intern · with the Mayo Clinic
B.Tech, Computer Science & Engineering · GPA 3.55/4.00