I built a lightweight CNN image-restoration pipeline during Intel Unnati Industrial Training 2025, using knowledge distillation on 3,214 GoPro blurred-sharp image pairs. The model reached 0.87 SSIM after 100 training epochs.
I combined four loss components in a PyTorch encoder-decoder and reduced training loss by 58%, from 0.0567 to 0.0237. I also transferred pretrained Restormer knowledge into StudentNet to support compact edge inference.
At the Geospatial Data Science Lab at GITAM University, I analyzed water-quality data across 21 chemical and biological parameters. I used stratified validation, StandardScaler, SMOTE, and six classifiers, with Random Forest reaching 95% accuracy and XGBoost reaching 92% accuracy.
I also tackled a 1:7,739 class imbalance in the BoT-IoT dataset using SMOTE, ADASYN, and hybrid oversampling. My strongest model achieved 99.95% F1-score and 100% accuracy across evaluated classifiers.
