top of page
Click on a project title, or the image beside it, to view documentation.
Built a neural network entirely from scratch using only NumPy (no high-level frameworks) for binary classification of text documents containing Personally Identifiable Information (PII). Processed and vectorized synthetic PII data using GloVe 6B 300-dimensional embeddings. Implemented several best practices for stable and effective training, including He initialization, ReLU activations, Dropout, Weighted Binary Cross-Entropy loss to address class imbalance, and L2 regularization. Achieved 93.3% test accuracy, correctly identifying 21 of 22 masked PII documents and 7 of 8 unmasked PII documents.
Investigated the classic “chicken-or-egg” relationship between stock news sentiment and price movements using both FinBERT and VADER across the top 10 tickers by news volume. Processed nested news text, generated daily sentiment aggregates, and compared them to daily, intraday, and next-day returns over a one-year period. While Pearson correlations were consistently weak, Savitzky-Golay smoothed time series visualizations revealed clearer patterns, suggesting bidirectional influence where news sentiment and stock prices appear to drive each other depending on the ticker — potentially reflecting differences between retail-driven and institutionally-driven stocks.
bottom of page