The various stages of an AI project flow

It’s been a little over4 weeks that we last communicated on the aspects of machine learning applications and the impact on the real world. In today’s blog, I would want to take off from where we left and talk a bit around the various stages involved in an AI project flow including machine learning model development, deployment and the monitoring of the models.

Let us start with the steps involved in the ML model development. We should then also briefly discuss the various data science tools used in these various stages of the ML Ops life cycle. Before getting into the ML Ops life cycle, I want to start with the example of an AI project flow.

There are various challenges involved during this AI development process starting from the ideation of a use case to validation based on experimental data, realization of the use case by development of the AI functions, productization by embedding these AI functions into business processes and finally operationalization aspects where these AI embedded processes are delivered to the customers. At each stage, the challenges are handled by adapting the required AI development methodologies.

This brings us to the topic of Embedded AI and a quick re-cap into the market which helps us understand how these are segmented. The enterprise AI market is segmented into — AI platforms, Applications Embedded AI, Platform Embedded AI. In each of these categories, the competition is intense (Microsoft Azure, Amazon AWS, Google GCP etc.,) and particularly when we talk about AI platforms or the platform Embedded AI, the players are many since there is no stickiness with regard to the applications! Applications embedded AI is a different beast which is where many of the business software firms like SAP, Salesforce, Oracle etc., have an edge! A holistic AI strategy consists of using an AI factory and building AI-supporting functions and distributing them into the AI eco system. The concept of an AI factory is typically a collaboration model between the applications and technologies in building embedded AI applications.

Let us now look into the various steps involved in an AI project flow in the context of building generic AI functions:

• Ideation — Use case identification based on common domain and AI expertise.
• Validation — Experiments to assess feasibility.
• Realization — Development of AI functions.
• Productization — Integration of AI functions into the business processes.
• Operations — Deliver embedded AI functions to customers and partners (the eco system).
• Continuous Improvement — Incrementally improve AI functions.

As you might see here, AI technologies are combined with the knowledge about business processes and integrated into the business applications. A typical AI factory model covers all steps of an efficient application development from ideation to productization in the evolution of a true enterprise grade AI application.

Let us now dig deep into these various steps in the context of ML ops life cycle and also understand the various frameworks used in each of these steps as well from onboarding to model serving. In the Ideation and the validation phase which involves data collection, data validation and feature engineering we can use frameworks such as the Sci-Kit Learn, PyTorch, TensorFlow, Keras etc. In the realization, productization and operations phase we can use tools like the JupyterLab, Metaflow libraries etc. The tools mentioned above are independent and the models developed can easily be deployed onto any enterprise embedded AI applications.

As we saw, the AI functions built using any frameworks are cloud agnostic and can run on any cloud provider. The following picture explains a typical end-to-end process flow with the different steps involved. This is a basic flow that I am putting here for the purpose of understanding — some of the steps might differ while doing the real development.

AI project flow

Now that we discussed about a typical AI project flow, let us now get into the aspects of an ML Ops life cycle and the steps involved in an end-to-end process:

• Onboarding:
• Service subscription and Service Key creation,
• Service subscription and role assignment,
• Setting up GitHub Repository.
• Configuration:
• Docker registry,
• Set-up API Environment for API calls,
• Object store registration API to create access secrets for Object Store,
• Creating Resource group to isolate the ML workloads,
• Training:
• Create Workflow template on the ML workbench leveraging the Argo Workflows,
• Register training dataset as an artifact,
• Configuration to be created which defines input artifacts and parameters to run training execution,
• Training executable which is an Argo workflow template that takes the input parameters and registered input artifacts like training datasets,
• Finally training execution which is an instance of a training executable that produces a trained model as an output artifact.
• Serving:
• Serving template on the workbench which is used to create model servers,
• Register trained model as an input artifact,
• Configuration to be created which defines input artifacts and parameters to run a deployment,
• Serving executable is a serving template that defines how a model should be deployed,
• Deployment is an instance of a serving executable that is configured to use the trained model as an input artifact and once the deployment URL is generated to be used for inferencing the model.
• Manage:
• Finally managing AI operations, AI services and the content,
• Monitoring the ML models for performance, data, algorithms, tuning etc.

As you can see, all these steps explained above constitute the ML Ops lifecycle and can be handled with any of the open-source frameworks and integrated into the enterprise embedded AI applications. With this basic explanation, we can now get into the next aspects of what type of tools are used in the AI model development. The following picture explains some of the data science tools that are used in the AI development and is given here to provide some basic understanding in the AI development process.

Data science tools in AI development

In the earlier section while discussing the steps involved in an end-to-end process of an AI project flow, we went into the technical details and hence let us now focus on the business aspects of the use cases:

• Data exploration and labelling — understand use case and the business process with exploratory data analysis, data labelling, basic experiments with simple models.
• Data cleaning and preprocessing — transform raw data into input data for machine learning model.
• Model training and validation — train machine learning model and evaluate business metrics, produce trained model artifacts.
• Inference and deployment — serve trained model and integrate with business process, track quality metrics during deployment.

I had a detailed podcast conversation with a couple of SAP experts in the context of AI project flow and their customer experiences which is available here for a leisure listening!

I would conclude this blog with these business aspects of the AI project flow and in the next blog let us get into the finer details of a complete ML Ops lifecycle.

Until next time, stay tuned and happy predicting the future with AI technologies!