Written .
MLOps deals with operationalizing machine learning models. This means taking models after their research stage and making them accessible via application programming interfaces (APIs) or HyperMedia. Difficulties seem to arise concerning the application binary interfaces (ABIs), data types and remote procedure calls (RPCs).
For example, layering different models seems fraught with variability in technical implementation. Consider object detection in an image and subsequent textual description in different languages. Chaining different models is currently a one-shot endeavour. This in turn, seems to hamper usage and value extraction.
Furthermore, tracking the history and training data of different versions of models is also a bespoke process. It is done manually.
Today's meeting was an opportunity to explore the practicality of different approaches. It seemed like a good environment for learning.
Shoutout to Garrett, Ethan and Joseph.