AI · Document AutomationProject 08
AI-Powered OCR Pipeline
Converts unstructured PDFs and itineraries into database-ready JSON. Combines OpenAI, Mistral AI, AWS, Azure, MinerU and OCR, syncing to Directus — cutting manual entry 80%+.
[ overview ]
A multi-format document-analysis engine that benchmarks cloud OCR services on real documents — PDFs, Excel, Word, PowerPoint, CSV and plain text — then scores extraction quality and maps exactly where each engine fails. Built as a production proof-of-concept at 7Span to choose the most accurate, lowest-cost OCR stack for automating manual data entry.
[ what it does ]
- Dual-engine comparison — Azure Document Intelligence vs MinerU Cloud on identical inputs
- Multi-format parsing (PDF, Excel, Word, PowerPoint, CSV, TXT) with automatic format detection
- Document-structure detection: multi-column reordering, invoice grouping, table preservation
- Accuracy metrics — similarity %, character/word accuracy, CER and WER (Levenshtein)
- Character and word confusion maps that show precisely how each engine misreads
- Dual-format baseline output (TXT + Markdown) plus structured JSON reports per document