Debanna Das

Advanced Cardiovascular Imaging Lab, University of Pennsylvania

SUPERVISOR: Dr. Walter Witschey

I lead Penn's hemothorax automation effort, training a 3D nnU-Net from scratch with 5-fold cross-validation, median-voxel resampling, and z-score intensity normalization to stabilize trauma CT volumes. This bespoke model lifts mean Dice from the TotalSegmentator baseline of 0.64 to 0.70 (Δ +0.06), and a fine-tuned variant, initialized with pleural-effusion weights and tuned via adjusted optimization parameters, lands at 0.68 (Δ +0.04). This clinically meaningful improvement has the potential to reduce radiologist workload and accelerate patient diagnosis, demonstrating how machine learning addresses real clinical challenges while maintaining scientific rigor and deployability.

Xu Lab, Carnegie Mellon University

SUPERVISOR: Dr. Min Xu

I developed a self-supervised Cryo-ET discovery pipeline that pairs a 3D SimCLR encoder with UMAP + HDBSCAN to recover label-free subtomogram motifs, moving silhouette scores from 0.16 to 0.43. The work demonstrated that dense molecular structure can emerge from unlabeled data when representation learning and manifold clustering are co-designed for volumetric biology.

Mit Media Lab

Optimized the Self-Organizing Network of Aggregated Representations (SONAR) stack by refactoring critical services, adding automated linting, and streamlining deployment workflows. The work reduced code complexity, hardened CI/CD, and ensured the federated perception system stayed maintainable as contributions scaled.

Genie Ai

Designed a modular AI system architecture with optimized data pipelines and workflow automation that improved processing efficiency and supported high concurrent user loads. Focused on making evaluation, observability, and deployment loops as rigorous as the underlying models.

Routechef

Built a multimodal travel assistant atop LangGraph and LLMs, integrating multiple travel APIs, Redis caching, and WhatsApp/web delivery channels; also shipped ML-powered chat analytics that reached 92% accuracy and improved user engagement.