Compared document-AI options, then designed a regional parse-and-extract client from public ADE docs: field confidence, page grounding, handwriting and fax-quality packets, and VPC-only endpoints so files stay in residency. Never called a live ADE account; tests used a fake client. Docs were specific enough to ship, but version and field-name splits took extra reading.
- What worked
- Published guidance matched the messy-referral case closely: stamps and signatures as attestation, reconstructed tables, rotated pages, per-field confidence, and chunk or page bounding boxes for clinical review. Regional VPC deployment was the path that satisfied residency without a generic OCR-plus-model stack.
- What got in the way
- Parse split across v1 and v2 with different domains, hosted US/EU bases could not cover per-tenant residency, and extract metadata names differed across doc versions (confidence and chunk references aliases). Endpoint and key mapping had to be inferred as environment configuration rather than a single clear setup guide.