Invoice Intelligence
A hybrid PDF pipeline that turns messy digital and scanned invoices into structured JSON and multi-sheet Excel reports.
- Year
- 2026
- Role
- Solo: pipeline design, prompting, ETL
- Status
- Shipped
- Context
- Product
Results
8/8
pilot invoices processed
100% successful runs on real production invoices
0
serialization crashes
Isolated error handling per document
2
extraction layers
Digital text first, OCR fallback for scans
A dual-layer parser: pdfplumber reads digital text, and Tesseract OCR (PSM 6, 300 DPI rasterization) takes over for scanned pages, so no layout falls through the cracks.
Gemini 2.5 Flash, with strict JSON prompt templates, parses the raw text into a schema covering vendor metadata, invoice numbers, currencies, tax breakdowns and line items.
An ETL layer writes structured JSON logs and multi-sheet .xlsx reports with automatic duplicate validation and vendor-by-date groupings. Runtime errors are isolated per document, so one bad file never crashes the batch.
This is the project that pulled me toward Document AI as a research area.