S0ME2/voight ? reverse-engineered prompt

Build me a FastAPI service for OCR that can handle three specific document types, an Uzbekistan passport, an Uzbekistan ID card with front and back together, and a driving licence.

It should accept an uploaded image and return structured extracted data with confidence scores, using document localization, image cleanup, region filtering, OCR, parsing, and validation. I want a ready to run API with a health check, Swagger docs, and example routes for each document type.

Please make it work well in CPU only development and Docker, with the model assets prepared during the image build. Include a clean project structure, config driven document profiles, tests for the core flows, and the usual developer commands for install, test, and run. If you need to, look up current docs for the OCR and FastAPI pieces online before wiring it together.

Also include the supporting docs and benchmark or validation helpers so the project feels complete, but keep it focused on these three supported layouts rather than trying to solve general OCR.