Route uncertain invoice fields to a human before using the extracted data
At a glance
SyncAI's founder adds human verification to an invoice extraction workflow that previously relied on OCR and language models. This workflow is listed under Document Extraction, Human in the Loop, Workflow Integration.
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Prepare a backend service, an invoice PDF or image, and an HTTPS webhook endpoint. Sign in to SyncAI's dashboard, generate an API key, and keep it in the server environment. 2) POST to https://sync-ai.up.railway.app/webhooks/register with the X-API-Key header and JSON containing url, events [document.verified, document.rejected], and secret. Use your HTTPS callback and chosen HMAC secret. 3) POST the invoice as multipart field file to https://sync-ai.up.railway.app/v1/extract using the same authentication header. Read status and document_id. 4) Use extracted_json when status is verified. If it is needs_review, wait for the callback. Verify X-SyncAI-Signature against an HMAC-SHA256 of the raw body using your secret; handle duplicate document IDs and acknowledge valid callbacks with HTTP 200. 5) On a verified event, pass its data downstream; keep rejected or unresolved records out of that completed path.
Worth knowing
The launch post and documentation disagree on the confidence threshold, citing 99% and 95%. Branch on returned status rather than assuming a cutoff. No independently verified accuracy is established.