Bayesian nonparametric approaches to causal inference have recently become popular. However, current approaches fail to address three important features in applications: Accounting for multilevel structure, allowing for targeted regularization, incorporating flexible models of the error distribution, and providing interpretable summaries of scientifically meaningful quantities. We extend recently proposed BART-based methods (Hahn et. al. 2017) to include all of these features. A key feature of this model is a parameterization that allows treatment heterogeneity to be regularized separately from prognostic effects, and also parsimoniously incorporates multilevel structure. In an application to the National Study of Learning Mindsets (Yeager et. al., 2017), a randomized experiment conducted in a nationally representative sample of schools, we use these new tools to provide meaningful insights about effect modification at the school and individual level. Our novel posterior summarization strategy avoids pitfalls common to existing approaches relying on post-hoc data snooping, model specification search, and/or large collections of hypothesis tests.