ML/AI expertise is not the problem
(Really! It's not.)
The importance and complexity of domain
knowledge is highly underrated
Industry ML is not a Kaggle competition but a
fit of Statistics/ML to an industry problem
with a lot of boundaries (data, systems,
Data that is not there cannot be magically
conjured (There's less/worse data than one
might think... always)
Specialized platforms might work nicely for a
case but I'd have to have very good reasons
to make our IT cloud infrastructure more
complex for just one case
So ... wisdom of the day: Scalability and
simplicity over niche solutions and (minor)
improvements in model quality.
Questions to ask before investing
> where do you get your training and test data from?
> do you obtain it legally?
> what is the quality of this data?
> what is the recall of your classification/prediction?
> what is the precision of your classification/prediction?
> what is more important for your customers? recall OR precision
> do you need to understand your model (legal & compliance) or not?
(When yes, you cannot use neural networks)