FlowSmith · sample deliverable
Five invoices, including an EU-formatted one, an ambiguous double-total and a bad OCR scan. The pipeline only auto-approves when its weakest field clears 0.80; everything else goes to a human. Accuracy below is measured against hand-labelled truth, not self-graded.
| file | vendor | invoice # | date | ccy | total | confidence | status | reason |
|---|---|---|---|---|---|---|---|---|
inv_1001_clean.txt | Northstar Studio LLC | INV-2026-1001 | 2026-10-28 | USD | $1,627.50 | 0.90 | auto-approved | — |
inv_1002_eu.txt | Alpen Data GmbH | RE-2026-0442 | 2026-10-30 | EUR | €2,023.00 | 0.90 | auto-approved | — |
inv_1003_ambiguous.txt | Acme Supply Co | INV-2026-1188 | 2026-11-02 | USD | $880.00 | 0.50 | → review queue | currency inferred from symbol, not an ISO code; conflicting totals on the document (840.00, 880.00) -> human review |
inv_1004_ocr.txt | Novak C0ntent | INV-2026-2010 | 2026-11-03 | USD | $1,050.00 | 0.55 | → review queue | currency inferred from symbol, not an ISO code; vendor name contains OCR artifacts (digits) |
inv_1005_clean2.txt | Ruiz Legal | RL-8842 | 2026-11-04 | USD | $750.00 | 0.85 | auto-approved | — |
approved.csv -> 3 row(s) to accounting review_queue.csv -> 2 row(s) to a human (Slack / sheet / email)
Text extraction (or OCR) → LLM returns a fixed JSON schema → confidence per field → review queue → approved rows written to their accounting system. Includes a 20-document accuracy baseline, a hard monthly API cost cap, error alerts and a 10-minute handover video. This demo uses rules instead of an LLM so it runs offline; the UX and review routing are the same.