* in the future it will use contextual bandits to probe for broken backups (such as: hey, is file XYZ with size X uploaded at Y weird?) and learn from all the c

* in the future it will use contextual bandits to probe for broken backups (such as: hey, is file XYZ with size X uploaded at Y weird?) and learn from all the customers triggering the AI flywheel, more backups more data, better contextual bandits, better service, more customers (the idea is to use bandits[or something else with exploration factor] for probability of "good" notification) i think the problem needs exploration, i have been bitten by bad anomaly detection and setting up alerts post-moretm after losing data one too many times :) i imagine creating a bunch of features such as: * file extension * time of upload * delta from previous version * size difference from other files in the directory * etc.. and having enough customers we should be able to do good-ish prediction when a file version is "weird", so we could send notifications such as: > hey is this file ok? with some exploration factor, even using UCB will(should) be better than nothing I plan to use vowpal wabbit's contextual bandits with only 2 actions, "send notification"/"dont send notification" given all the context this of course will be just extra on top of the manual alert rules, but hopefully it will save some data :) If the users agree we could also publish anonymized datasets with labeled data (such as: given context alert was sent: was [good]/[bad]) Which will be awesome.