Our summary
The final lesson in Joe Herbert’s Maia and Jev series starts with a bucket of emails, call transcripts, portal submissions and scanned claim forms. He asks Maia to build a pipeline that reads and cleans these files, asks Jev questions about each claim and turns the results into organised warehouse tables.
A file registry records what needs processing. Questions live in a separate table, so changing them does not require rebuilding the pipeline. Jev’s answers are stored beside confidence and token counts. Herbert corrects the design to keep each claim, rather than each claimant, at its centre, then demonstrates tests, release to a checking environment and traces of where data came from.
Herbert reports processing 440 claims, but has not published accuracy against the answer key. The build does not reconcile claims with a central database or enforce a confidence cutoff. Its estimated seven-cent evaluation bill comes from earlier measurements. The reported ninety-minute build also reused earlier components. Treat this as an inspectable demonstration, not validated claims automation.
Key takeaways
- Separate reading mixed documents from asking AI questions. A file registry helps reveal inputs that were missed rather than silently processing fewer files.
- Keep questions in a table and answers linked to each claim. Store confidence and cost information so reviewers can inspect individual results.
- Test decisions against labelled claims and add a review route before using them to act. A data trace or confidence column does not establish correctness.