LLM Document Data Extraction with Schema Validation & OCR

I need an engineer to build/harden a document data-extraction pipeline — not a chatbot, not a Q&A assistant. The system reads long, often scanned documents (PDFs with tables, multi-column layouts, numbered clauses) and must output a verified, structured checklist/dataset (JSON) that other systems depend on. Must have: Structured/schema-enforced LLM output (Pydantic, Instructor, or equivalent function-calling approach) — no free-text parsing Real experience with messy PDFs: scanned pages, tables, multi-column layouts, OCR when needed A concrete method to verify extracted fields against the source text (verbatim/fuzzy matching, confidence scoring) and flag/escalate low-confidence extractions to a human reviewer — zero silent misses is the requirement Experience measuring extraction quality (test/golden set, precision-recall, or similar) — not just "it works on my examples" Cost-aware production setup (batching, caching) for processing documents at volume