I think prediction APIs are a joke at this time.<p>The assumption here is that the hard part is running an SVM or a deep network or random forest or something like that.<p>The hard part is really defining the problem, cleaning the data, figuring out the features and parameter tuning the system. This involves building models many times, and if (i) the clock is running on the algo, and (ii) you don't know enough to implement the algo, you stay safely on the wrong side of the commercialization valley of death.