Applied ML and LLM systems, built one project at a time

Predicting appliance energy usage from indoor/outdoor sensor data. Four models compared, from linear regression to Random Forest, with full diagnostic analysis.

Classification under extreme class imbalance (0.17% fraud rate), with SMOTE/class weighting comparison, threshold tuning, and SHAP explainability.

Unsupervised network intrusion detection on NSL-KDD. Isolation Forest vs One-Class SVM, with analysis of detection on never-before-seen attack types.

Time series forecasting compared across ARIMA, Prophet, and Random Forest — including why two structurally different models failed in the exact same way.

Unsupervised customer segmentation via RFM analysis. K-Means clustering cross-validated with hierarchical clustering, visualized with PCA.

CNN for automated visual inspection of cast metal parts — plus a data leakage audit uncovering why the model scored a suspicious 100%.

Transfer learning with a frozen ResNet18 — training only 0.046% of the network to classify satellite imagery into 10 land use categories.

Comparing Whisper model sizes on clean vs. challenging audio — Word Error Rate weighed against inference time, not accuracy alone.

Transfer learning on Mel spectrograms to classify emotion in speech — plus a reproducibility bug caught and fixed before it could bias the result.

A hand-built Canny/Hough pipeline against a transfer-learned U-Net — including a diagnosed structural failure on Botts' Dots reflectors and a reworked evaluation metric to compare a line against a mask fairly.

DistilBERT fine-tuned to sort CVE descriptions into CWE categories — plus two rounds of error analysis, one uncovering real label noise, the other correctly leaving a lookalike confusion alone.

A system-prompt-engineered study assistant, documented naive vs. refined — plus three real Claude Sonnet 5 API behaviors discovered and worked around along the way.

Retrieval-augmented generation over three arXiv papers, with every answer traced to a numbered citation — plus a cross-document retrieval limitation found, tested at higher top-K, and documented rather than hidden.