You can build things very quickly with Edge Impulse, but it allows you to go deep into a problem as well. Here’s me going over a typical anomaly detection usecase where you have very little examples of anomalies. It covers building custom DSP blocks (Python & C++), determining important features, testing out your model, and finally on implementing your custom blocks on device:
(This was shot back in December, so if I reference stuff that was not public yet - like feature importance - that’s the reason 
)