Alex Galvez-Vega

Owlivia is a retrieval-augmented AI advising assistant I have developed for Florida Atlantic University's Electrical Engineering and Computer Science department since May 2025. It uses a documentation database built from curated public department content to answer questions about courses, forms, schedules, and academic requirements.

I built the retrieval pipeline around BM25 sparse search and FAISS dense-vector search, then merged the ranked results with reciprocal-rank fusion before supplying grounded context to the language model. I integrated the system into the chatbot backend, tuned retrieval and model behavior, and deployed it with web access to improve answer coverage while keeping department documentation as the primary source.

This project recreates the lightweight ADD-Net convolutional neural network for classifying Alzheimer's disease stages from MRI scans. The work examines the model architecture, class imbalance, training behavior, and interpretability.

The recreation compares the original SMOTETOMEK balancing approach with alternative sampling methods including SMOTEENN, ADASYN, and Borderline-SMOTE. Grad-CAM visualizations provide interpretable heat maps showing the image regions that influence model predictions.

The hybrid-search system runs BM25 keyword retrieval and FAISS dense-vector retrieval in parallel. BM25 prioritizes exact terminology, while the embedding search retrieves passages with similar meaning even when the wording differs.

Reciprocal-rank fusion combines the two ranked result lists into one retrieval set. The resulting passages can then supply grounded context to a retrieval-augmented generation pipeline.