State-of-the-art hot stamping processes offer improved material formability and therefore have potential to successfully form challenging components. The feasibility of components to be formed through these processes is dependent on their geometric design and its complex interactions with the hot stamping environment. In industrial practice, trial-and-error approaches are currently used to update non-feasible designs where simulation runs are needed each time a design change is made. These approaches make the design process resource intensive and require considerable numerical and process expertise. To demonstrate a superior approach, this study presents a novel application of a deep-learning-based optimisation platform which adopts a non-parametric geometric modelling strategy. Here, deep drawn corner geometries from different geometry subclasses were optimised to minimise wasted volume due to radii while avoiding excessive post-stamping thinning. A neural network was trained to generate families of deep drawn corner geometries where each geometry was conditioned on an input latent vector. Another neural network was trained to predict the thinning distributions obtained from forming these geometries through a hot stamping process. Guided by these distributions, the latent vector, and therefore geometry, was iteratively updated by a new gradient-based optimisation technique. Overall, it is demonstrated that the platform is capable of optimising geometries, irrespective of complexity, subject to imposed post-stamped thinning constraints.