Intelli
Docs

The challenge

A team of forty people were retyping documents into a database. Invoices, contracts, claims forms — hundreds of layouts, arriving as PDFs, scans, and phone photos taken at an angle in bad light. Off-the-shelf OCR handled the clean documents and quietly mangled the rest.

The real risk was not slow processing. It was a wrong number entering a financial system with nothing flagging it.

What we did

We designed the pipeline around one principle: the system must know when it is unsure. Accuracy on its own is a vanity metric if the failures are invisible.

• A preprocessing stage that de-skews, denoises, and normalises everything from crisp PDFs to phone photos.
• A layout-aware classifier that identifies document type before extraction, so each type gets the right model.
• Field-level extraction with a calibrated confidence score attached to every value.
• A human-in-the-loop review queue that surfaces only low-confidence fields — and feeds corrections back as training data.
• An event-driven architecture that scales horizontally through volume spikes without a queue backing up.

What shipped

A platform that ingests documents from email, SFTP, and API, and emits structured, validated records into downstream systems. Anything below the confidence threshold is routed to a reviewer with the source document and the uncertain field highlighted side by side; every correction improves the next batch.

Results

The forty-person data entry team became a small review team handling exceptions. Nothing reaches the financial system unvalidated, and the error path is visible rather than silent.

Working together

We started with the client's messiest documents, not their cleanest, because a pipeline proven on the worst inputs needs no caveats later. Weekly demos ran against a live sample from production. More on how we scope this kind of work in the FAQ.