aiptimizer/TurboOCR ? reverse-engineered prompt

Build me a very fast self hosted document OCR service for Linux that uses an NVIDIA GPU and can run in Docker. I want it to handle both images and PDFs, return plain OCR text with boxes and confidence, and also support richer document parsing like layout detection, reading order, tables as HTML, formulas as LaTeX, and full page or PDF output as Markdown when requested.

Please expose simple HTTP endpoints and a gRPC service, with sensible defaults for speed, plus optional settings to trade speed for better accuracy on harder scans. It should work well for forms, receipts, and dense documents, support the built in model size options and the extra language recognizers mentioned in the docs, and report which capabilities are currently loaded. I also want Prometheus style metrics and a clean Docker setup where TensorRT engines are built on first start and cached so later restarts are fast.

If anything is unclear, check the current project docs online and wire it up end to end so I can run it and test it right away.