Debanna Das

Advanced Cardiovascular Imaging Lab, University of Pennsylvania

SUPERVISOR: Prof. Walter Witschey

Hemothorax—internal chest bleeding from trauma—requires rapid diagnosis, but manual CT analysis is slow and subject to human variability. I developed an AI-assisted detection system at Penn Medicine to help radiologists identify these critical bleeds more quickly and consistently. The system combines multiple deep learning architectures working together: 2D models analyze slice-by-slice patterns while 3D models capture full volumetric context. By merging their predictions, this ensemble approach outperformed both individual models and fine-tuned versions of existing medical imaging systems. The work shows that architectural diversity, not just more data, can drive better clinical AI performance. This reduces diagnostic delays in emergency settings and gives radiologists a reliable AI assistant for detecting these life-threatening bleeds.

Xu Lab, Carnegie Mellon University

SUPERVISOR: Prof. Min Xu

Cryo-electron tomography generates millions of 3D images of molecular structures inside cells, but manually labeling these volumes is prohibitively slow, which creates a bottleneck for biological discovery. I developed and compared two unsupervised discovery pipelines to automatically identify and group similar molecular structures without requiring human annotations. The first approach used transfer learning: a pretrained 3D ResNet-18 from video recognition extracted volumetric features, followed by t-SNE for dimensionality reduction and K-Means clustering to discover structural groups. The second approach trained a custom 3D SimCLR encoder from scratch using contrastive self-supervised learning on the cryo-ET data itself, then applied UMAP for manifold learning and HDBSCAN for density-based clustering that automatically handles noise and determines cluster counts. The self-supervised approach outperformed the transfer learning approach, nearly tripling cluster coherence and producing substantially better-separated molecular groupings. This enables high-throughput structural discovery in areas where expert annotation is the rate-limiting factor for scientific progress.

MIT Media Lab

Contributed to SONAR (Self-Organizing Network of Aggregated Representations) at MIT Media Lab, a decentralized federated learning framework enabling peer-to-peer collaborative model training across distributed nodes without centralizing raw data. Refactored core algorithm modules and utility services to enforce consistent coding standards, introducing static analysis tooling via Pylint and Pyright with enforced score thresholds and pre-commit hooks to catch regressions at the source. Hardened the CI/CD pipeline by authoring GitHub Actions workflows covering automated linting, type checking, and functional end-to-end training tests over the gRPC communication backend — ensuring new contributions remained reliable as the codebase scaled across both MPI and gRPC deployment environments.

GENIE AI

At GENIE AI, I worked across the full ML development lifecycle—from studying LLM architecture fundamentals (transformer attention mechanisms, positional encodings, and layer normalization strategies) to fine-tuning GPT models for domain-specific applications in healthcare diagnostics and educational content generation. Developed systematic evaluation frameworks that benchmarked fine-tuned outputs against ground-truth behaviors across precision, recall, and domain-specific accuracy metrics. Led the system architecture design for GenCanvas, defining the end-to-end data flow: user natural language inputs processed through spaCy NLP pipelines for entity extraction, routed to domain-specific GPT endpoints with custom system prompts, validated against JSON schemas, and rendered as interactive canvas elements via real-time WebSocket connections.