Totally agree that it's too slow with conventional approaches, which is why we're training custom models for this that we can run fast