Built an AI-powered fashion styling demo using Python, Streamlit, FashionCLIP, FAISS, and Qwen2.5. The app extracts style, color, occasion, and fit preferences from natural-language prompts, then uses semantic search over fashion-image embeddings to recommend coordinated outfits across five categories: outerwear, tops, bottoms, shoes, and bags. Includes alternative recommendations and an interactive web interface.

Developed a handwritten character recognition application using a Convolutional Neural Network (CNN) trained on the EMNIST dataset. Built with PyTorch and Streamlit, the app lets users draw a character on an interactive canvas and receive a model prediction. The project explores image preprocessing, deep learning-based classification, and real-time model inference through a user-friendly interface. The model achieved approximately 86.6% accuracy on the test set.

Developed a customer churn prediction and risk analytics project using Python, SQL, and machine learning on a 10,000-record banking dataset stored in SQLite. Tuned the classification threshold to 0.2697 to support churn-risk identification, flagging 2,433 customers as high risk. Built an interactive dashboard to explore customer risk and communicate insights that can support data-driven retention strategies.

Developed a multimodal medical AI research prototype combining chest X-ray images and medical text representations using PyTorch, DenseNet121, and BioBERT. Implemented contrastive learning to explore cross-modal representation alignment and Grad-CAM to visualize image regions influencing model outputs. The project explores computer vision and biomedical NLP for medical image-text analysis. This is an experimental research prototype, not a clinically validated diagnostic tool.