OCR in-process
Scanned PDFs read without sending documents to a third party, on the API hosts themselves.
Built and operated by us
Scanned paper into a searchable, self-filing archive
Uploads and scans arrive, text is extracted or OCR'd in-process, keywords and labels are found, and a rules engine files, renames, labels and expires documents on triggers. Every rule can be previewed before it runs, and conflicting rules are caught.
Document capture fails in two ways: the OCR is wrong and nobody notices, or the automation does something nobody expected. We built against both: every stage is versioned so it reruns when improved, and every rule shows what it would do before it does it.
The same parts make an AP invoice or a bill of lading readable and routable in a distributor's back office, which is the service this case study underwrites.
What we built
A pipeline and a rules engine.
Versioned stages
Extraction, OCR and labelling each carry a version; improving one reruns only that stage.
Rules with twelve conditions
Text, keyword, file name, type, folder, size, dates, label and OCR state, with six actions from label to delete-after-N-days.
Preview before it acts
Any rule can be tested against the current documents and explained for one document.
Conflict detection
Rules that fight each other are found before they run.
History and revert
What a rule did, when, and the way back.
Why it matters to you
What this proves.
How does this become AP automation for us?
The same pipeline reads a supplier invoice; the rules match it to a purchase order and route it to an approver; the archive keeps it. The missing piece is your ERP's PO and receipt data, which is an integration we also do.
Does our paper leave our building?
Not unless you want it to. OCR runs in-process on the hosts we control, and nothing is sent to an outside AI service unless that trade is worth it to you.
Who checks the machine?
A person, always. The preview and explain tools exist so the person sees what the machine saw before anything is posted.
Count the documents someone typed last week.
Thirty minutes, no obligation. We reply within one business day.