Case studies

Challenge

Traditional federated learning systems struggle with non-IID data distributions and lack interpretability, making them difficult to deploy in real-world scenarios where data privacy and model transparency are critical.

Solution

Developed a novel explainability-driven framework that integrates attention mechanisms and explainability techniques to improve robustness and efficiency under non-IID conditions. The framework uses spatial attention layers (SPATL) and explainability constraints (XLC) to enhance model performance and interpretability.

Technologies

Results

• •Published in IEEE Access (2025)
• •Improved model accuracy by 15% on non-IID datasets
• •Enhanced interpretability through attention visualization
• •Reduced communication overhead by 20%

Impact

This research contributes to making federated learning more practical for privacy-sensitive applications in healthcare and finance.

Challenge

Understanding and implementing state-of-the-art generative models from scratch to gain deep insights into diffusion processes and denoising mechanisms.

Solution

Built a complete denoising diffusion probabilistic model (DDPM) implementation using PyTorch, including U-Net architecture, time embeddings, EMA (Exponential Moving Average), and gradient clipping. Created a full training pipeline with visualization tools.

Technologies

Results

• •Successfully generated high-quality images from noise
• •Implemented complete training pipeline from scratch
• •Created interactive demo using Streamlit
• •Achieved stable training with EMA and gradient clipping

Impact

Demonstrated deep understanding of generative AI and provided educational resource for others learning diffusion models.

Challenge

Deep learning models often have slow inference times, limiting their deployment in production environments, especially on edge devices or high-traffic applications.

Solution

Developed a comprehensive optimization pipeline comparing multiple techniques: torch.compile, TorchScript, ONNX Runtime, and mixed precision inference. Built FastAPI inference services with Dockerized deployments for easy benchmarking.

Technologies

Results

• •Achieved 3x speedup using torch.compile
• •Reduced model size by 50% with quantization
• •Created production-ready inference API
• •Documented optimization trade-offs and best practices

Impact

Provides practical guidance for ML engineers optimizing models for production deployment.

Challenge

Build a scalable HR management system with responsive UI, robust backend, and clean architecture principles to handle employee data, payroll, and performance tracking.

Solution

Developed a full-stack application using React for the frontend and Node.js with Express for the backend. Implemented RESTful APIs, MongoDB for data storage, and responsive design with Tailwind CSS. Applied clean architecture principles for maintainability.

Technologies

Results

• •Delivered scalable HR management solution
• •Implemented responsive design for all devices
• •Built robust API with error handling
• •Applied clean architecture for maintainability

Impact

Streamlined HR operations and improved data management for the organization.