This paper presents Ubervvald, a comprehensive library designed to enhance classification tasks by optimising complex Convolutional Neural Networks (CNNs). It integrates advanced techniques for improving accuracy and computational efficiency. Additionally, it is designed to be modular, allowing one to either configure a custom optimization pipeline, reutilizing its functions, or just use the default pipeline by a mere 2–4 function calls. Ubervvald achieves robust performance in various real-world scenarios by leveraging multi-task learning and cascaded network architectures. Key features include automated hyper-parameter optimizations, gradient-based optimization in parameter numbers, types and model size reduction for edge device deployment while providing a high degree of flexibility target device-wise. The proposed solutions within Ubervvald demonstrate significant improvements over existing methods, making the library a valuable tool for researchers and developers in computer vision.