TALK ABSTRACT

Todays headlines are about AI/ML bias and the risks and associated horror stories about AI as an existential threat. However, it doesn’t matter if you think of ChatGPT as the first step toward the Terminator or the grandparent of C3PO - when you go beyond modeling, it is clear DevOps is critical to effectively implementing unbiased and accurate models in our enterprise and consumer applications.

In this presentation, I talk about how the DevOps cycle is become foundational to the success of AI/ML production transformation as more corporations begin development of apps and SaaS applications with AI at their core. I specifically cover: PLAN: Integrating AI/ML into the agile development and operating process BUILD: Eliminating bias, inconsistencies, non-compliance in data by going for Gold (a reference to the medallion data engineering process) BUILD/RELEASE: Enable repeatability and observability of AI models TEST: Support multi-environment E2E testing with real data and predictions MONITOR: Know the truth – the Ground Truth! MONITOR: Maintain, measure, and monitor “drift”

While the lessons here have been learned on multiple clouds in the deployment of models with different tool sets and languages, I am currently working with Azure DevOps, and with models deployed using Databricks using YAML and MLFlow, and the talk includes some specific references to ADO and Databricks processes in the slides.

Brad Taylor gives awesome advice on data and AI/ML strategy to global companies, private equity portcos and random strangers at AUSUM Advisors, and as a volunteer at Dallas AI, a 501c3 North Texas