The workflow before
Ford was moving hundreds of thousands of engineering requirements from Polarion and VSEM into JAMA. Each one had to be rewritten to meet INCOSE and EARS standards, which took an engineer about 10 minutes. A single batch held 13,035 requirements, roughly 2,170 hours of skilled work, before anyone checked that the meaning had survived.
What we built
We built an LLM pipeline on Ford's private Azure tenant that ingests raw requirements and rewrites each one to INCOSE standards. Models are closed-source and private, so proprietary engineering data never leaves Ford's environment.
Each rewrite is scored by the JAMA Advisor Connector. If the quality score is low, the pipeline loops back and refines the text until it passes; anything that still falls short goes to an engineer.
A second model runs in parallel and scores semantic similarity between the original and rewritten text, so a requirement cannot pass if its engineering meaning has drifted. A batch of 13,035 now rewrites and scores in about 4 hours.
Before and after
| Measure | Before | After |
|---|---|---|
| Time for a 13,035 requirement batch | About 2,170 engineer hours | About 4 hours (2.5 rewrite, 1.5 scoring) |
| Time per requirement | About 10 minutes | About 1 second |
| Meaning drift check | Manual peer review | Similarity score on every rewrite |
What's next
The same rewrite, score and refine loop applies to the rest of the requirements backlog and to new requirements as engineers write them, catching standard violations at authoring time instead of during migration.
