Rizky Ananda Putra

Classifies Indonesian-language health claims as myth or fact, deployed as a real-time Telegram bot. The 200-entry seed dataset was validated with a certified medical practitioner, then expanded via back-translation. Cross-validation accuracy reached 97.5% (weighted F1 97.4%) — though accuracy drops to 55% on out-of-distribution external data, a limitation documented as a direction for future work on word embeddings.

An end-to-end TabNet pipeline — an attention-based deep learning architecture for tabular data — on a 3,235-record clinical dataset with IQR outlier handling. Validated via Stratified 5-Fold Cross Validation: 70.9% accuracy, 71.1% precision, 75.8% recall on average, with no signs of overfitting. Deployed as a Streamlit app for real-time risk prediction.

A web-based decision support system scoring doctors across 4 criteria and 12 sub-criteria (administrative, performance, violations, external factors) to recommend license renewal or revocation. Weight consistency verified across every pairwise comparison matrix (Consistency Ratio ≤ 0.10).

Maps 33 symptom parameters to 4 disorders — Insomnia, Sleep Apnea, Narcolepsy, Restless Legs Syndrome — using Forward Chaining (symptoms → diagnosis) and Backward Chaining (hypothesis verification), scored with Certainty Factor theory. Example outputs: CF 0.94 for Insomnia via Forward Chaining, CF 0.97 for Restless Legs Syndrome via Backward Chaining.

Unsupervised clustering of patients by treatment priority across a 16-feature medical dataset. Optimal cluster count (k=3) chosen via the Elbow Method, with cluster separation confirmed by Silhouette Score above 0.5. PCA used for 2D visualization of the resulting groups.

Classifies cat vocalizations into 14 categories (13 emotional states + non-cat noise) using SVM. In-memory signal augmentation — noise, pitch, and speed perturbation — expanded 566 raw recordings into 1,700 training samples to correct severe class imbalance, reaching 92.94% overall accuracy. Presented at STAINS 2026.

An interactive app benchmarking three classic classifiers across 3 Kaggle datasets — binary classification on fruit and pumpkin varieties, multi-class on 9 fish species — with Grid Search tuning per algorithm-dataset pair. Best results: 96.5% test accuracy (fruit), 94.0% CV score (fish), 89.0% CV score (pumpkin).