Experience.

Developing a two-prior score-based diffusion model that removes structured clutter from transesophageal echocardiography, training unpaired tissue and clutter priors on 13K patches from 58 patients.

Building an inverse finite element pipeline to predict calcified nodule stiffness from CT volumes, replacing assumed density-to-stiffness values with a transferable CT-to-stiffness law for TAVR planning.

Client: Boston Public Library (Digital Commonwealth).

Architected a hybrid retrieval pipeline over 1.8M+ chunks of the Digital Commonwealth archive, fusing BGE-M3 dense and sparse retrieval with reciprocal rank fusion to raise Hit@50 from 0.40 to 0.66.

Added a Neo4j GraphRAG layer with spaCy entity extraction and GPT-4o cited generation, deployed as an app that grounds every answer in real archival sources.

Built BLIP ITM and SigLIP 2 scoring pipelines to filter noisy frame-caption pairs from a 264K-clip infant touch dataset, preparing clean data for multimodal contrastive pretraining.

Benchmarked vision-language scorers against human-annotated labels to set calibrated filtering thresholds, contributing to a dataset paper accepted at IEEE ICDL 2026.

Delivered a CAD framework for an ICMR-funded study classifying benign versus metastatic pulmonary nodules in sarcoma patients, trained on 3D CT from 460 patients with patient-wise splits.

Designed an ensemble of multi-scale 3D CNNs with spatial and channel attention reaching 0.85 sensitivity and 0.85 AUC-ROC, with Grad-CAM++ maps to surface model reasoning.