Production-document inspection was manual and repetitive. In an FDA-regulated operation, those checks still had to be done carefully. I built a workflow that automated routine inspection and routed exceptions to people for review.
Start with the process and the documents
I redesigned the inspection process with the team and revised the document templates for more consistent extraction. The templates were part of the engineering: both reviewers and models needed clear inputs before I could automate the checks.
Give the model a specific job
I used vision-based extraction to turn production records into structured data. Prewritten validation scripts handled repeatable checks, and agents used controlled tool calls to work through the inspection.
The workflow combined:
- Document templates designed for consistent extraction
- Rule-driven validation alongside model output
- Controlled access to the tools agents needed
- Explicit routing of edge cases and exceptions to human reviewers
Review the exceptions and improve the workflow
The agents automated nearly all routine inspection, leaving people to review exceptions. I used evaluation results to refine the prompts, skills, and rules as the workflow developed.
This project reflects how I approach AI in an operation: understand the process, decide which checks belong in code, and make it clear where a person needs to take over.