This study presents an inline quality inspection system for resistance spot welds that predicts nugget diameter and classifies surface defects using 3D scanning and machine learning. Indentation geometry data from a high-resolution 3D scanner are processed to extract 16 geometric features, analyzed by a hybrid model combining a convolutional neural network (CNN) and a rule-based algorithm. Surface defects are classified into six types: normal weld, edge weld, electrode tip degradation (ED), tilting, spatter, and pinhole, with ED divided into No ED, Low ED, and High ED based on its impact on weld quality. In inline operation, dual robots and a movable tracker cover the full body-in-white (BIW) within cycle time. The workflow integrates scanning, feature extraction, nugget prediction, and defect classification for immediate OK/NG decisions. A 3D visualization interface displays BIW components with color-coded weld status, process parameters, defect types, and ED levels, enabling rapid feedback. Validation with 62 specimens across various sheet combinations and field variables achieved high accuracy in both nugget prediction and defect classification, with an average inspection speed of 2.5 seconds per weld. This approach overcomes the limits of destructive and conventional non-destructive tests, offering a reliable inline solution for real-time feedback and smart factory implementation.