Modern make-to-order (MTO) job shops are vulnerable to disruptions, ranging from machine failures to supply chain fluctuations, threatening their ability to meet delivery commitments. This study introduces a framework designed to enhance operational resilience by integrating the Relative Lateness Proportion (RLP) to quantify delay propagation, along with the Active Period Percentage (APP) for monitoring productive capacity. These metrics are combined with an advanced Attention-based Long Short-Term Memory (A-LSTM) prediction model. Validated through extensive testing in a real-world precision manufacturing facility, the proposed solution demonstrates significant advantages over conventional methods, including Autoregressive Integrated Moving Average (ARIMA), vanilla Recurrent Neural Network (RNN), and basic LSTM architecture. The framework enables the early identification of vulnerable workstations, accurate prediction of disruption patterns, and data-driven interventions to maintain production schedules. Its performance is attributed to integrating operational metrics with machine learning, where the attention mechanism effectively prioritizes critical production data. The proposed approach provides manufacturers with a practical and scalable method to improving delivery reliability in complex, high-variability production environments.