Machine learning surrogate for large-scale wave energy converter arrays using geometric feature engineering to predict farm-level absorbed power from array layouts, published in the Journal of Ocean Engineering and Marine Energy (Springer Nature).

Residual SwiGLU Deep MLP for Higgs boson signal identification, achieving 88.52% ROC-AUC, 89.50% PR-AUC, and 79.99% accuracy on the full 11M-row HIGGS dataset, after benchmarking LightGBM, XGBoost, autoencoder-enhanced trees, and physics-engineered feature baselines.

BERT-based 150-class intent classifier achieving 96.31% accuracy, supported by a full few-shot evaluation pipeline (10–50 samples per intent) using deduplicated, leakage-free splits. The project also focuses on a complete Out-of-Scope detection system built with Outlier Exposure, energy-based scoring, and Mahalanobis distance—achieving 0.98 AUROC in distinguishing in-domain from unseen queries.

BERT-based 7-class intent classifier achieving 98.56% accuracy, and a DistilBERT slot-filling model reaching 94.5% entity-level F1 on the SNIPS 2018 dataset. Few-shot experiments reached 97.84% accuracy with only 70 samples per intent. The full pipeline includes preprocessing, leakage-safe splits, tokenization, training, and evaluation.

Lightweight phishing URL detection using pre-fetch features only, with incremental learners reaching 96.19% accuracy.

95.3% accuracy on the 3-class fetal health (NSP) task, and 91.7% accuracy on the 10-class fetal heart rate (FHR) classification problem.

88.29% AUC and 80.88% accuracy for gallstone detection using 10-fold cross-validation.

Retained ~99% of full-model ROC-AUC using only 50% of the original features via permutation-based feature ranking and K-fold cross-validation.