Amazing post! This is why Monzo is light years ahead. Keep these types of posts coming.
I’d be interested to hear how long this experimentation process took approximately from start to finish. We have a similar technical problem which we believe will be solvable with Machine Learning but I’m not sure what a reasonable time to expect to come to a decent solution is. Thanks.
I started at Monzo mid Feb 2017, and one of main tasks was to work on augmenting customer support (displaying relevant answers, suggesting tags etc.). I think I spent around a third of my time working on this. The help search model uses one of the models used in customer support augmentation, so work on the help search itself was pretty quick. It was the initial customer support problem that took way longer, and is still a work in progress.
Experimentation took a while, but the bulk of the time was spent on research - reading papers, exploring methodologies etc. If your problem has a similar structure to our’s, feel free to use the same kind of training methodology we used (using triplets in training + a hinge loss function). We found that the most effective.