Projects

We design a two-branch architecture: a CNN backbone (ResNet-50) for local texture and edge features, and a ViT module for global relational reasoning. Features are fused via a learned attention gate before classification.

Achieved competitive sensitivity and specificity compared to standalone CNN and ViT baselines on the test split. Grad-CAM visualizations show attention focused on lung fields consistent with TB pathology.

We evaluate 12 augmentation strategies (geometric, photometric, mixup-based) across DenseNet-121, EfficientNet-B3, and ResNet-50. Performance measured with AUC, sensitivity, specificity, and F1-score.

CutMix combined with geometric augmentation yielded the best generalization, improving AUC by ~3% over no-augmentation baseline on the held-out test set.

Transfer learning with MobileNetV3 and EfficientNet-B0. Applied class-weighted loss to handle dataset imbalance. Evaluated with k-fold cross-validation.

Achieved over 93% accuracy on 4-class classification with consistent performance across cross-validation folds. Grad-CAM overlays highlight relevant kidney regions.

Built a full-stack platform with Next.js frontend and Go backend, integrating Semantic Scholar API and OpenAI API for summarization. Supports tagging, citation tracking, and Markdown note-taking.

MVP deployed and tested by a small group of graduate students. Positive feedback on literature search and paper summarization features.

Fine-tuned YOLOv8 on a curated dataset of emergency event frames. Built a REST API with FastAPI, containerized with Docker, supporting stream input and event webhook callbacks.

Achieved real-time inference at 30+ FPS on GPU. Deployed and tested on a local surveillance testbed with 94% precision on the evaluation set.