Machine Learning Experiments.
Comprehensive suite of practical machine learning implementations covering Deep Learning Computer Vision (CNNs), Natural Language Processing (BiLSTM Sentiment Analysis), and Unsupervised Anomaly Clustering (PCA & K-Means).
Computer Vision (Natural Scene Classification)
An end-to-end Deep Learning pipeline for natural scene image classification across 6 categories (Buildings, Forest, Glacier, Mountain, Sea, Street) trained on the Intel Image Dataset.
Architecture & Deployment: 4 Convolutional blocks, ReLU activations, MaxPooling, Flatten, Dense layer (512 neurons), 0.3 Dropout regularization, and Softmax classification. Exported to TensorFlow SavedModel for production servers, TensorFlow Lite for mobile edge, and TensorFlow.js for in-browser client inference.
Natural Language Processing (Gojek Review Sentiment)
Complete NLP workflow analyzing user sentiment on the Gojek application from Google Play Store reviews, classifying feedback into Positive, Neutral, and Negative categories.
Pipeline: Custom review scraping, Indonesian text cleaning, tokenization, word embeddings, Bidirectional LSTM layers, Dense ReLU layers, and Softmax classification. Monitored with confusion matrices and classification reports.
Unsupervised Anomaly Detection (Fraud Transaction Clustering)
Unsupervised customer transaction anomaly detection grouping financial behaviors to isolate suspicious activities and potential fraudulent patterns without labeled training data.
Methodology: Multi-feature scaling (Transaction Amount, Age, Duration, Login Attempts, Balance), correlation matrix analysis, Principal Component Analysis for dimensionality reduction, and K-Means clustering.
Open Source Training Repositories
image-classification-tensorflow
CNN natural scene classifier with 88% accuracy, SavedModel, TFLite, and TF.js exports.
sentiment-analysis-gojek
Indonesian review sentiment classifier with Bidirectional LSTM and 88.69% accuracy.
fraud_detection_clustering
Unsupervised customer transaction clustering with K-Means and PCA visualization.
fraud-detection-classification
Supervised classification models for predicting high-risk transaction patterns.
Eksperimen_SML_Miqdad
Supervised machine learning benchmarking suite across tabular prediction datasets.
Workflow-CI
Continuous integration pipeline automating model training checks and deployment tests.