Supplier unpredictability remains a key challenge for small-scale Vietnamese companies with limited technological adoption. This study analyzes one year of engine import data from an automotive manufacturer to predict late deliveries using machine learning classification models. To address data scarcity, CTGAN was used to generate synthetic data for enhanced model training and testing. Five models: Random Forest, Decision Tree, XGBoost, LightGBM, and Logistic Regression were evaluated using three key preprocessing steps: (1) SMOTE balancing, (2) feature augmentation, and (3) mutual information-based feature selection. On the SMOTE-balanced real dataset, Decision Tree outperformed other models with the highest Accuracy (0.82) and lowest false positive rate (0.2708), followed by XGBoost with strong balanced performance (Accuracy of 0.80, FPR of 0.2917). On synthetic dataset, LightGBM (Accuracy of 0.81 and FPR of 0.2997) also demonstrated strong predictive power. Compared to the company’s baseline, where 80% of deliveries were late and no predictive model was in place, this ML pipeline significantly improves forecasting reliability, enabling proactive intervention and planning. Although results are not yet production-ready, this study offers a strong foundation for future adoption in SMEs.. The study also offers valuable insights for small and mid-sized manufacturers globally, particularly in emerging markets with constrained access to quality data. Future work should explore alternative synthetic techniques (e.g., CopulaGAN, Synthpop-NP) and investigate CART models for improved data quality and prediction scores.