Projects

Scaled-up successor to mahaMiniGPT for training a SOTA decoder-only Transformer from first principles. Integrates modern techniques like Flash Attention, SwiGLU, RMSNorm, and GQA for efficient pre-training. A full-stack, research-grade framework for reproducibility and understanding of modern LLMs.

Efficient unlearning framework with error-maximizing noise generation. Fast unlearning without full retraining while maintaining optimal performance. Designed for GDPR compliance and model privacy preservation.

An end-to-end implementation of modern Vision Transformer (ViT) architectures like DeiT, Swin, and DINO for image captioning. Focuses on comparing different visual encoders with a Transformer-based text decoder on benchmark datasets like COCO and Flickr30k.

Systematic exploration of LoRA hyperparameters for fine-tuning the Qwen 2.5 7B model. Detailed ablation studies on PEFT and LoRA settings for efficient, scalable Python code generation. Achieved up to 93.34% execution success rate using the flytech/python-codes-25k dataset.

Data-free class impressions framework for selective model forgetting. Achieves ≈0% accuracy on forget classes without original training data while maintaining performance. Proven effectiveness from MNIST to ImageNet across LeNet5 to ResNet50.

Strategic game analysis and mathematical modeling using game theoretic principles. Exploration of Nash equilibrium, strategic interactions, and comprehensive game simulations.

Deep learning-based image inpainting for content restoration and reconstruction. Leverages DDPM and advanced architectures for natural and seamless image completion.

An educational project to demystify the Transformer architecture by building it step-by-step. Empirically measures the impact of each component—from basic attention to modern optimizations. Serves as a "living research notebook" to provide data-driven answers to core architectural questions.

A collection of repositories including BasicML and DeepLearn that implement fundamental and Deep Learning algorithms from scratch. Focuses on building a strong foundational understanding using primarily Python and NumPy.

A foundational repository implementing classic Machine Learning algorithms from the ground up using Python and NumPy. Focuses on understanding the core mechanics of models like Linear Regression, K-Means, and more.

Final Year Project focused on detection of road potholes using computer vision and deep learning. Analyzes real-time images/video feeds to identify and locate potholes on urban and rural roads. Includes severity scoring and geo-tagging for actionable road maintenance insights.

IoT-based security system for intruder detection and surveillance using Raspberry Pi. Upgraded from Arduino version with PIR sensor integration for motion detection. Captures images/videos on motion trigger and sends email notifications with visual evidence. Features camera module, email alerting system, and real-time monitoring capabilities.

Secure password management application with strong encryption and auto-lock functionality. Implements AES-256 encryption for offline password storage with SHA-256 hashing. Features intelligent auto-lock on inactivity, password generator, and hierarchical organization.