MLOps: Creating powerful AI pipelines by stitching together heterogeneous Machine Learning models

As Machine Learning and Deep Learning are being widely adopted in different industries, in many cases there is also a need to use these models together to have a system or a solution that is capable of performing much better on a task than the individual models can. Or even perform tasks that the individual models cannot perform by themselves in isolation. We developed the Acumos AI Platform that provides capabilities to create powerful AI pipelines by stitching together heterogeneous Machine Learning models. We will demonstrate with multiple examples how we can create such powerful AI pipelines which we will call ‘composite ML solutions’ using the Acumos Design Studio. We will use some other components we developed i.e. Acumos Runtime Orchestrator, proto viewer, Data broker, Splitter, and Collator to deploy the aforementioned composite ML solutions as a set of communicating microservices. The individual ML models that are part of the composite solution may not necessarily have been created by the same person or same group or the same organization or have been created in a particular language or use a particular ML framework. But still, Acumos allows these models created in disparate languages and frameworks by different individuals or organizations to be stitched (subject to their compatibility) together to create composite solutions using a simple intuitive GUI and deploy them very easily to various cloud targets. In this paper, we will describe these novel capabilities that we have developed in Acumos and we will show how to create and deploy a composite ML solution or an AI pipeline using this toolset.