Reducing defects and creating new industrial products are issues addressed in literature, though often separately. This work presents a unique approach combining machine learning and optimization techniques, which could 1) predict and explain ongoing production defects, 2) prescribe solutions to prevent predicted defects, and 3) produce defect predictors for novel products without historical data. By applying it into a melamine-surfaced boards process, task 1) explored SVM, XGBoost and Random Forest algorithms, coupled with an explainable model-agnostic approach. A Powell’s method-based meta-heuristic algorithm handled task 2), while the Hyper-Process Model (HPM) technique was selected in task 3). Defect prediction presented strong results, with a fine-tuned XGBoost with oversampling achieving over 0.8 recall value, and prescriptions significantly reduced prediction scores across most defect scenarios. Although the HPM implementation introduced challenges producing predictors, most generated synthetic data-trained models achieved encouraging performances on recall metrics. Results could serve as baseline for future developments.