Variational Inference (e.g. Variational Bayes) can use a variety of approximating densities. Some recent work has explored using classes of Generative Neural Networks with Jacobians that are either volume preserving or fast to calculate. In this work we explore two points: using more general neural networks, but taking advantage of the conditional density structure that arises naturally in a Hierarchical Bayesian model and a general inference framework, in the Spirit of David Spiegelhalter’s WinBugs software, where a wide range of models can be specified and the software ‘automatically’ generates an approximation of the posterior density.