Our work demonstrates an innovative AI-driven approach to Quality Assurance (QA) in metal additive manufacturing (AM). We focus on real-time defect detection in 3D metal printing, which is increasingly used to manufacture small parts for applications in safety-critical devices, including parts for the aerospace, automotive, and medical industries. We propose a model-based failure and defect detection system that combines visual monitoring and input from various sensors to control the AM process. By using a deep learning model and incorporating live monitoring of data points, we enable early anomaly detection and thus ensure immediate operator intervention to stop and prevent defective prints to be used in safety-critical applications. These techniques help prevent energy waste, conserve material costs, resources, time, and increase the safety and reliability of the system. Our prototype framework automates the classification of multiple defects, improving both the reliability of printed components and the efficiency of the manufacturing process by suggesting in-situ improvement techniques. We also highlight the technical challenges associated with data synchronisation and acquisition, as well as multiple sensor integration. Data collection is used not only for QA but also as valuable documentation for process validation and to maintaining an in-house learning center. Overall, our AI-driven approach improves the accuracy of non-destructive testing (NDT) over traditional post-production inspection methods. We propose a framework for the integration of AI to improve process control, making it more reliable for safety-critical industries and providing strong guidance for future model-based QA applications in AM processes. This framework can be applied for QA on TRUMPF TruPrint L-PBF 3D printing machines.