At VVDN Technologies, I conducted a theoretical comparison of YOLOv5s and YOLOX-Tiny, deriving their label-assignment mathematics and loss formulations. I assessed accuracy scenarios for dense clustering, small-object detection, and extreme aspect ratios against AP/GFLOPs tradeoffs.
For an Institute Technical Summer Project, I developed an offline face-recognition attendance system using HOG detection and dlib encodings. It included automated enrollment, a Flask REST API, and a dashboard with live streaming.
In my image-classification projects, I reached 90.3% accuracy on Fashion-MNIST with an ANN and improved it to 93.8% by fine-tuning VGG 16. For surface crack detection, I fine-tuned ResNet50 and achieved 99.48% accuracy on unseen data, with results shown in a React dashboard.
My polymer-informatics work applied Adaptive Conformal Prediction to XGBoost, RF, and SVR models, validating 90.4% coverage at 90% confidence. I also fine-tuned ViT-Base for steel microstructure classification, outperforming ResNet50 by 6 points.

