Integrating Machine Learning Models and an AI Agent for Decision Support in Engineer-to-Order Manufacturing

Abstract

Engineer-to-order manufacturers must plan projects under substantial uncertainty: operation durations are difficult to estimate, shop-floor capacity is constrained, procurement delays can reshape feasible schedules, and external material-market disruptions can alter execution risk. This thesis develops and evaluates an integrated decision-support framework on real industrial data. The framework combines three predictive and planning components: a two-stage model for manufacturingorder duration prediction, a discrete-time shop-floor simulator that converts duration predictions into dependency- and procurement-aware project plans, and nickel and copper disruption-forecasting pipelines formulated around alert episodes. These components are orchestrated by a tool-using large language model agent, built on a locally hostable open-weight model, which grounds numerical claims in tool outputs and translates uncertainty into natural language. On held-out data, the duration pipeline predicts realised order durations with a mean absolute error of 1.15 working days (R2 = 0.71). The commodity pipelines produce disruption alerts that improve on a momentum benchmark for both metals. An informal evaluation with managers found the agent's combined outputs clear and actionable, with trust and capability scope emerging as the main reservations. These results indicate that engineer-to-order decision support becomes more operationally useful when prediction, interpretation, and cross-model reasoning are treated as a single design problem rather than as separate tasks.