Our summary
The second lesson in Joe Herbert’s Maia and Jev series puts five kinds of work through the shared integration. It sorts support tickets, rates content, checks policy documents, studies product reviews and labels possible customer churn. The examples show how to store decisions in a data warehouse, not how accurately they predict real outcomes.
Choice selects a category, Score rates something on an ordered scale and Noul estimates the probability of a yes-or-no condition. Several questions can share the same information. Herbert also adds a documentation team to ticket routing, reviews Maia’s proposed changes and reruns the pipeline without building a separate integration.
The inputs are small and synthetic. The customer-churn labels have not been checked against later customer behaviour. Herbert reports costs for forty requests, then projects larger bills using published prices. Those comparisons are not matched accuracy tests. Keep arithmetic and date checks in code, allow none-of-these answers and evaluate ambiguous real examples before relying on the decisions.
Key takeaways
- Use Choice for a category, Score for an ordered rating and Noul for a yes-or-no probability. Store their different answer shapes deliberately.
- Change the questions while reusing the pipeline, then test the updated categories. A fixed answer list needs a way to handle inputs that do not belong.
- Separate a working integration from proof of accuracy. Compare Jev with the cheapest alternative that already meets your task’s quality needs.