Awesome - INLA like idea is an area we would love to have more implementation - similar paper in case you have not seen it is: [2004.12550] Hamiltonian Monte Carlo using an adjoint-differentiated Laplace approximation: Bayesian inference for latent Gaussian models and beyond