Client work (NDA)
OCR + AI document analysis
Purchase invoices and delivery documents into the ERP
90% less manual data entry across hundreds of documents a day, with a review queue instead of quietly passing errors through.
Context
A company re-keying purchase invoices and delivery documents into internal systems by hand (shown anonymized under NDA). Built before KSeF, when a PDF was the only input a supplier ever sent.
The problem
Hundreds of documents a day re-keyed by hand. The expensive part was not the typing but the missing control over what got keyed in wrong: a bad amount or the wrong counterparty surfaced in accounting weeks later, once it was costly to unwind.
What we built
Documents land in a dedicated inbox. An Airflow pipeline classifies the document type, cleans the scan with OpenCV, runs Google OCR, and an LLM turns the result into structured line items. Then the checks: line items against the document total, counterparty tax ID against the contractor database, duplicate detection. Every field carries a confidence score, and that score drives the branch: above the threshold the document goes straight into the target system, below it lands in a review queue, and anomalies alert over e-mail and Slack. We treat the threshold as a business decision rather than a technical setting, because raising it means fewer errors and a longer queue while lowering it means more automation and more risk. Success was measured with two numbers that only work together: the share of documents clearing the pipeline untouched, and the share the system waved through despite an error. The first one is only allowed to grow while the second stays at zero. Containerised and idempotent, so reprocessing the same document never doubles anything.
Results
- ▹90% less manual data entry
- ▹Hundreds of documents a day in steady operation
- ▹Low-confidence extractions queued for review instead of passing silently
- ▹Cross-validation of totals, tax IDs and duplicates, with e-mail and Slack alerts