Comment by jliszka

Yep. We normalize our features as part of training, and the stdevs of each feature are part of the resulting model, along with the weights. (The means are always 0 because of the way we construct our training set.) The weights we use in production are actually normalized_weight / stdev.