Invoice chaos → clean ledger.
A real intake pipeline, engine room exposed: pick a messy invoice, run it, and watch extraction, validation and posting happen in seconds. One of them has a problem — see if the pipeline catches it before you do.
The document~8 min to handle by hand
INVOICE
Meridian Analytics Ltd.
14 Clerkenwell Road, London EC1M 5PA
VAT Reg No: GB 345 6789 01
Invoice no.: INV-2026-0847
Issue date: 14 Sep 2026
Due date: 14 Oct 2026
| Description | Qty | Unit | Amount |
|---|---|---|---|
| Data pipeline maintenance (September) | 1 | 1,850.00 | 1,850.00 |
| Dashboard seats | 12 | 35.00 | 420.00 |
| Ad-hoc analyst support (hours) | 6 | 95.00 | 570.00 |
IBAN GB29NWBK60161331926819
Please quote the invoice number on payment.
The pipeline
- 1Read documentOCR text layer
- 2AI extractionClaude · structured output
- 3Validationdeterministic checks
- 4Ledger entryor human review queue
This pipeline took a few days to build. Yours would plug into your actual inbox, ERP and approval flow — same idea, your systems.
How it works: extraction runs server-side with Claude using a strict output schema; validation is deterministic code, not AI. When the live model isn't configured, the demo replays a previous run of the same pipeline — the badge above tells you which you got. Nothing you do here is stored.