Applied Projects
Breast Cancer Detection via CNN Architectures
Problem Statement
Does transfer learning meaningfully close the performance gap on a small, class-imbalanced medical imaging dataset?
Approach
- Compared custom CNN, fine-tuned VGG16, and ResNet50 on BUS-BRA (1,875 ultrasound images)
- U-Net segmentation to isolate tumour regions before classification
- SMOTE to address class imbalance; data augmentation for generalisation
My Contribution
- Designed and ran all three pipelines end-to-end
- Implemented the U-Net preprocessing stage independently
- Ran statistical comparison and authored the findings
Outcome
ResNet50 achieved 93.4% vs custom CNN's 85.2% confirming that transfer learning compensates effectively for limited data in this domain. Results submitted for peer review.
Women's Health Recommendation System
Problem Statement
Generic health recommendation tools ignore intra-cycle hormonal variation, what a woman needs nutritionally and physically in the luteal phase differs significantly from the follicular phase. Can an LLM pipeline produce phase-specific, evidence-backed guidance at query time?
Approach
- RAG pipeline for evidence retrieval at inference time
- Google Flan-T5 base for initial generation; FastAPI backend
- Phase classification from user inputs, mapped to retrieval queries
My Contribution
- Designed the full RAG architecture and Gemini API integration
- Built and deployed the FastAPI backend
- Evaluated recommendation quality across all four cycle phases
Outcome
RAG-augmented pipeline outperformed baseline generative approach on query resolution efficiency. Work accepted for publication in Innovations in Healthcare Technology and Management.
Tools & Technologies
Languages
ML & Deep Learning
Dev Tools