Sheet stamping is a widely adopted manufacturing technique for producing complex structural components with high stiffness-to-weight ratios. However, designing such components is a non-trivial task that requires careful consideration of manufacturing constraints to avoid introducing defects in the final product. To address this challenge, this research introduces a novel deep-learning-based platform that optimises 3D component designs by considering manufacturing capabilities. This platform was realised by developing a methodology to combine two neural networks that handle non-parametric geometry representations, namely a geometry generator based on Signed Distance Fields (SDFs) and an image-based manufacturability surrogate model. This combination enables the optimisation of complex geometries that can be represented using various parameterisation schemes. The optimisation approach implemented in the platform utilises gradient-based techniques to update the inputs to the geometry generator based on manufacturability information from the surrogate model. The platform is demonstrated using two geometry classes, Corners and Bulkheads, each having three geometry subclasses, with four diverse case studies conducted to optimise these geometries under post-stamped thinning constraints. The case studies demonstrate how the platform enables free morphing of complex geometries, while also guiding manufacturability-driven geometric changes in a direction that leads to significant improvements in component quality. For instance, one of the cases shows that optimising the complex component geometry can reduce the maximum thinning from 45% to satisfy the thinning constraint of 10%. By utilising the proposed platform, designers can identify optimal component geometries that ensure manufacturing feasibility for sheet stamping, reducing design development time and design costs.