Many optimal insurance design problems with the incentive-compatible (IC) condition lack closed-form solutions and efficient numerical methods. This talk proposes a novel algorithm to obtain numerical solutions of optimal insurance by parameterizing the indemnity function using a Lipschitz Multilayer Perceptron (MLP) architecture. To incorporate the principle of indemnity and the IC condition, we introduce constrained affine transformation layers and a special activation function while preserving the MLP’s universal approximation. We provide a theoretical justification for the proposed architecture and develop a gradient-based algorithm that efficiently approximates optimal indemnity functions with very high accuracy. We validate the algorithm’s accuracy against classical benchmarks with known analytical solutions, and demonstrate its applicability to several economically important yet previously intractable settings including optimal insurance with a dependent background risk, loss ambiguity, or rank dependent utility. A case study on Florida hurricane insurance further illustrates how derived optimal contracts can mitigate the crowding-out effect of government relief to enhance welfare.