Our customer performs visual final inspection on safety-critical pivot pins with complex surface geometries. The components feature demanding structures such as knurled areas and multiple functional surfaces, all of which require highly precise image capture. The previous solution was based on a rule-based machine vision system. In practice, this resulted in a high false reject rate because fluctuations in image quality, lighting, and part positioning had a major impact on inspection results. At the same time, the defect classes were too complex for rigid rule-based logic—especially in cases where identical defects had to be evaluated differently depending on their position on the part. Integration requirements were also demanding. Multiple robots, large camera distances, and camera positions that had to be both protected and precisely aligned made implementation particularly challenging. With Maddox AI, the existing rule set was replaced by tailored AI-based inspection models. These models learn complex, context-dependent defect patterns directly from image data. In addition, regions of interest can be defined for each image, making it possible to clearly distinguish between relevant and non-relevant areas and reliably map zone-specific defect catalogs.