Session Description

Generative AI can generate useful content at extraordinary speed, but the quality and reliability of that output remain uneven. Raw AI output can contain subtle errors, inconsistencies, weak explanations, or questionable engineering choices. Human review remains essential, but once AI can generate content at scale, human attention itself becomes the bottleneck.

This talk presents an engineering approach to improving the quality of generative AI output before humans enter the loop. Rather than treating AI as a standalone assistant, we can build “quality pipelines” around it using deterministic programs, repeatable workflows, structured review loops, iterative refinement, and layered human vetting.

As a case study, I’ll describe work done for an STA (Sovereign Tech Fund) Scala Center project that focused on improving the documentation of the Scala standard library. We wrote Scala programs that inserted TODO markers into .scala source files where Scaladoc was missing. We then used long-running Scala scripts to orchestrate an AI-driven workflow that generated documentation, reviewed generated text, produced structured feedback, and iteratively refined the results one file at a time before multiple stages of human review.

In this talk, you will learn practical approaches to building quality pipelines around generative AI, including deterministic orchestration, structured review loops, iterative refinement, and layered human vetting. The session will focus on applying traditional software-engineering ideas to AI systems to achieve the quality you want while making the most effective use of the precious resource of human attention.