Techniques for adaptive pipelining composition for machine learning (ML)

This invention describes an automated system for continuously improving machine learning applications. It works in the background to identify new machine learning models or better settings (parameters) for existing models. These new models or settings are first tested offline against real-world data. If they perform better than what's currently in use, the system automatically upgrades the live application with the improved version, sometimes after a 'shadow mode' period where it runs alongside the old one for validation.