I built a personal project for explanation-aware chest X-ray classification, combining DenseNet-121 with a Grad-CAM loss on VinDr-CXR. The training penalizes attention outside expert-annotated disease regions.
I also implemented an end-to-end pipeline for DICOM preprocessing, training, and evaluation, and created custom interpretability metrics. A FastAPI service returns predictions with Grad-CAM overlays.
For a separate personal project, I implemented SimCLR from scratch in PyTorch and pretrained it on unlabeled STL-10 images. Its linear probe achieved 76.7% test accuracy, compared with 78.7% for a fully supervised ImageNet-pretrained baseline.
At New Line Software Company, I developed RESTful APIs and Laravel backend services integrated with MySQL, and wrote unit and integration tests. Earlier, as a Software Engineering Intern at Trillium Soft Software Company, I implemented responsive mobile interfaces, state management, and REST API integration.
