A Systematic Study of Unsupervised Domain Adaptation on Sparse Educational Data

• Evaluated CORAL for Domain Adaptation across 42 transfer pairs (N=32,593)
• Identified 18.0% accuracy drop caused by feature structure destruction in sparse data
• Utilized SHAP to expose the "Illusion of Fairness" in misleading fairness metrics

Architected edge-first AI system for rural healthcare settings

• MobileNet V3 + TensorFlow.js with Dual-Axis Projection Analysis
• Digitized smudged medication instructions and decoded indigenous handwriting patterns
• Achieved 100% offline reliability in low-connectivity environments
• Multi-Signal Matching Architecture with 1024-dimensional visual embeddings

Privacy-preserving, blockchain-based aid distribution system

• Built using Polygon ID and Coinbase SDK
• Enabled secure and decentralized aid distribution
• Protected beneficiary privacy through zero-knowledge proofs

Implemented 6 end-to-end PyTorch projects in computer vision

• Real-time object detection with YOLOv8
• Face recognition using MTCNN
• Medical image generation with GANs
• Deployed interactive vision applications using Streamlit and Flask

Developed e-learning platform for rational use of medicine

• 4-week e-learning and certification platform
• Promoting rational use of medicine in rural, low-connectivity communities
• Serving underserved communities across Nepal