Agile AI Platform Architecture with the Agile Cloud Manager
Part 5 of 10: Seed Each New AI Project With Standard Catalyst Templates
The entire transcript of this video is given below the video so that you can read and consume it at your own pace. We recommend that you both read and watch to make it easier to more completely grasp the material.
The number, the complexity, and the importance of A.I. projects add a lot of overhead to the amount of work that needs to be performed for each project.
Compliance with laws requires audit trails and security.
The risk of data drift and concept drift require that everything must be reproducible.
An effective way to meet all of the compliance and reproducibility requirements is to create and use standardized templates to create each new project.
The projects are defined using templating tools like cloud formation, ARM templates, or terraform templates.
Agile Cloud Manager can organize all those lower-level templates into an appliance that can be managed as a group and operated upon as a group.
So then you can use one single Agile Cloud Manager C.L.I. command to create each new project, complete with many different standardized components that we will describe in this slide.
- Document templates can be put into each new A.I. project, so that each of the types of work done in each project can be done more quickly, and at a higher level of quality that meets all of the requirements of your governance process.
- The directory structure for each of your A.I. projects can also be standardized so that it is easier for data scientists to switch between projects. For example, when creating new models to potentially replace existing models.
- The types of issues and work items for use in A.I. projects can also be defined at an organization level. That way, work within each A.I. project can be managed more effectively. Including examining how much energy is expended on each type of work.
- Workflow structures for A.I. projects can also be defined at the organization level.
- The same types of version-controlled repositories can be created for each project. For example: Data sets need to be version controlled for reproducibility. Models need to be stored in version control. Scoring code and A.P.I. wrapper code also need to be stored in version-controlled repositories.
- Environment definitions can also be deployed as a part of each new A.I. project.
A group in your organization can provide standardized baseline templates for environments that are defined as infrastructure as code, and as configuration as code. Regional variations for these standardized baselines can also be created in order to comply with local laws in certain parts of the world. And the software-defined environments can be deployed in a standardized sequence that is also baked into the templates that you define in order to create each new project. For example, each of your projects might define a dev environment, which might proceed to a test environment, and which would conclude with a production environment. - Pipeline structures can also be created each time you create a new A.I. project. A data pipeline will do all of the transformations required in order to turn raw data into a dataset that can be used for training models. A machine learning pipeline will manage all the work of building, training, and experimenting with A.I. models. And a DevOps pipeline manages the larger process of creating, provisioning, and maintaining the sequence of environments starting with dev and continuing through production.
- Role-based access control can be defined for every element of the predefined project structure. For example, data preparation might only be allowed by one type of person. And deployment to production might only be allowed by one different type of person.
- Event monitors can also be written into the templates that create each A.I. project. Some of the event types might be transitions between steps in project management workflows. Other types of events might involve changes in the substance of the results generated by monitors of production environment activity, for example when data drift is observed by automated testing, or when concept drift is observed by automated testing.
- Policies can also be defined as elements of the templates that are used to create each new A.I. project.
Altogether, the elements on this slide illustrate how software-defined project templates can enable your organization to more quickly iterate through A.I. projects while complying with laws and managing the unique life cycle of A.I. models, including data drift and concept drift.
Agile Cloud Manager makes it a lot easier to organize software-defined templates into coherent units that can be more easily managed, deployed, and iterated-upon.
Proceed to Part Six: Continuous Experimentation And Training
Back to Series Table Of Contents: Agile AI Platform Architecture With Agile Cloud Manager