The spam scale problem should be solvable with a one-shot locally evaluated NN trained to classify behavior/tweets from their massive massive ground truth dataset. If it’s evaluated locally then I don’t see how scale is an issue.