This thesis presents the design, development, and evaluation of a proprietary object detection model tailored for road surface defect identification, with a particular focus on potholes. Developed in collaboration with the Italian startup LOKI s.r.l. in relation to its project Asfalto Sicuro, the work addresses the limitations of relying on pre-trained third-party models by proposing fully customized deep learning architectures. Two pre-trained backbones —MobileNetV2 and Darknet-53— were each integrated into a YOLO-inspired architecture featuring a split detection head. Each model was trained and evaluated on a curated pothole detection dataset using extensive data augmentation techniques, including translation, cropping, mosaic patterns, and perspective distortion. Performance was benchmarked against state-of-the-art models (YOLOv8 and Faster R-CNN).