Actual recording · Synthetic English narration and captions · Fictional invoices
THE PROBLEM
Extracting fields does not complete an administrative process: data needs review, delivery, traceability and failure recovery.
THE WORKFLOW
Tesseract reads the image, rules validate the fields and a person corrects and approves. HTTP delivery reaches a separate SQLite-backed service with idempotency keys, a receipt and retries that preserve approved work.
01OCR and validation
02Human correction and approval
03Idempotent HTTP delivery
04Receipt, persistence and retry
RECORDED EVIDENCE
Results, with context.
Recorded checks on the stated dataset
Check
Cases
Fields correctly extracted before review
81 of 84
HTTP delivery and recovery checks
8 of 8
Dataset
document-relay-v1
Completed cases
20/20
Median / p95 per case
0.025 s / 0.031 s
Evaluation record (UTC)
2026-10-09
Source: local pipeline execution on the identified inputs. CPU; no GPU. HTTP timings refer to the interface samples; for reconstruction they measure loading recorded geometry, not a new reconstruction. These are not customer outcomes or a guarantee on other data.
A document workflow connected to the destination API, with review, traceability, deduplication and agreed maintenance.
What we would define first
Layouts and fields, human review, duplicate rules, destination API, retention, credentials and recovery.
Demonstration limits
Twelve fictional invoices of one layout: 81/84 correct fields before review. Eight integration checks passed over real HTTP. The receiving ledger is a persistent demonstration, not a customer ERP or bank. This does not measure general invoice extraction.
Recorded demo
Recorded demo