Large-Scale Land Survey Digitization Pipeline

A large volume of physical land survey records, roughly 3 million documents, existed only as scanned PDFs with no structured, searchable data behind them. Off-the-shelf OCR tools failed on the real scan quality: skewed pages, inconsistent lighting, and dense handwritten annotations that generic OCR benchmarks don't account for.

A fully open-source digitization pipeline that compresses, OCRs, and parses each document at scale. Vision-language models handle the documents generic OCR can't read, while a fine-tuned Tesseract instance handles high-volume standard text as the CPU workhorse, keeping cost manageable against a fixed project fee.

1. 01Documents are ingested and compressed for efficient processing and storage in Cloudflare R2
2. 02Preprocessing steps, deskew, denoise, CLAHE, adaptive threshold, are applied and tuned against real survey document scan quality
3. 03Each document is routed through the appropriate OCR engine: PaddleOCR-VL and Qwen2-VL-7B on GPU for complex or handwritten content, fine-tuned Tesseract on CPU for standard text

• →First successful end-to-end pipeline run confirmed on real survey documents
• →Preprocessing tuned specifically against real scan quality, not generic OCR test sets
• →Hybrid GPU/CPU OCR routing keeps processing cost sustainable at millions-of-documents scale

• →Handles ~3 million documents, a very different scale problem than a typical CRM or workflow build
• →Custom preprocessing tuned against real-world scan artifacts rather than clean benchmark data
• →Hybrid OCR strategy: vision-language models for hard cases, traditional OCR for volume, balancing accuracy against cost