This paper presents a novel deep learning-based platform for addressing shape distortion in sheet metal stamping (e.g., springback, thermal distortion) by tool compensation. Conventional approaches to tool compensation involve computationally expensive Finite Element (FE) simulations to update tool geometries. In contrast, the proposed platform uses a generator network to create 3D tool geometries and an evaluator network to predict the resulting shape distortion and post-stamping thinning. The generated tool geometries are iteratively updated by a gradient-based optimisation technique in the direction of minimising shape distortion in the resulting component. The platform is demonstrated on a cold stamped U-channel component case study, which experiences severe shape distortion in the form of springback. The optimisation problem was formulated to find the optimum tool geometry that enables a desired U-channel geometry to be formed after springback by tool compensation, while meeting a maximum thinning constraint. The platform successfully optimised the tool geometry to compensate for springback in this setting, showcasing its effectiveness in improving manufacturing outcomes and product quality. The presented approach offers a superior method for addressing shape distortion in stamping processes, as compared to conventional FE simulation iterations or trial-and-error methods. This approach can efficiently and effectively compensate for arbitrarily complex tool geometries without requiring extensive process expertise.