Distribution grids are currently challenged by frequent voltage excursionsinduced by intermittent solar generation. Smart inverters have been advocatedas a fast-responding means to regulate voltage and minimize ohmic losses. Sinceoptimal inverter coordination may be computationally challenging and presetlocal control rules are subpar, the approach of customized control rulesdesigned in a quasi-static fashion features as a golden middle. Departing fromaffine control rules, this work puts forth non-linear inverter controlpolicies. Drawing analogies to multi-task learning, reactive control is posedas a kernel-based regression task. Leveraging a linearized grid model and givenanticipated data scenarios, inverter rules are jointly designed at the feederlevel to minimize a convex combination of voltage deviations and ohmic lossesvia a linearly-constrained quadratic program. Numerical tests using real-worlddata on a benchmark feeder demonstrate that nonlinear control rules driven alsoby a few non-local readings can attain near-optimal performance.