Agentic Framework for Production Time Estimation of Sheet Metal Parts from Engineering Drawings Using Multi-Modal Document Analysis

Production time estimation in sheet metal manufacturing traditionally relies on expert knowledge and interpretation of engineering drawings, a process that is time-consuming and introduces variability in estimated times. Automating this task is challenging due to heterogeneous drawing styles for many small and medium companies, acting as suppliers, varying formats, and complex document structures that intermingle geometric, textual, and dimensional information. This paper presents a comparative study of techniques for automated feature extraction from mechanical drawings to enable machine learning based time prediction for laser-cutting. This paper proposes a multi-stage computer vision pipeline that sequentially filters drawing infrastructure like title blocks, removes dimensioning artifacts and identifies flat-view geometry for perimeter calculation. A combined approach is presented, where relevant features are extracted from drawings and used together with historical routing data to train regression models. Comparative experiments demonstrate that XGBoost achieves strong predictive performance (91.96\% accuracy, NRMSE = 0.0804) when trained on manually extracted features, validating the feasibility of the approach. However, automated extraction faces significant challenges, including format heterogeneity and scale detection ambiguity. This paper discusses these limitations and proposes future directions that incorporate YOLO-based object detection, convolutional neural networks (CNNs), and LangGraph orchestrated agentic frameworks with human-in-the-loop verification to improve extraction robustness and enable practical deployment.