About

Data scientists often spend months pursuing marginal improvements in model performance before a solution ever reaches production. The result is a familiar problem: no feedback, no adoption, and no business value.

This talk argues for a different approach. Using a real-world rostering system built at RAC, it demonstrates how deploying an imperfect solution early can create immediate operational benefits while creating the feedback needed to build a better product. Rather than treating deployment as the end of a project, production becomes part of the development process itself.

We will explore practical techniques for incremental deployment, risk management, stakeholder engagement, and learning from live operational data. It will be particularly relevant for data scientists, analysts, and technical leaders looking to shorten the path from model development to measurable impact for key projects.

Key Takeaways:

• How to select suitable projects and define a ‘good-enough’ threshold that justifies live deployment before a model is complete
• Monitoring and rollback strategies for models that are still being actively developed
• How early production data accelerates a model’s improvement faster than any offline evaluation