Dogfood AME on a non trivial project

[jmintb]

To test AME in a realistic setting before the first release we should build an ML project using AME. Ideally this should be done in a remote deployment in GCP with mixed CPU and GPU workflows. Implementing a simple image classifier based on stream recordings seems like the best option at the moment.

Acceptance criteria:

• There is a public repository using AME
• The project is non trivial
• The project uses the complete range of features AME as to offer.
• (Optional) The project can be displayed to others.