Fine-tuned BERT, RoBERTa, and AlBERTa using LoRA for efficient hate speech detection. Achieved over 90% accuracy with sub-110ms latency using encoder-based LLMs.

Developed a GA-based optimizer that outperformed standard methods like SGD and Adam by 2–5% on CIFAR-10. Integrated Optuna for efficient hyperparameter tuning.

Built a combined depth estimation and object detection pipeline, optimized via quantization for embedded devices. Achieved 0.801 mAP50 on real-time deployment with Raspberry Pi.

Designed a quantized DNN model achieving 99.92% accuracy and 90% size reduction. Integrated Grad-CAM to enhance explainability for real-world currency validation.

Performed feature-level ensemble learning and ADASYN-based augmentation to enhance Alzheimer’s classification. Compared multiple deep architectures for model performance.