The AI reads it.A person still signs it off.
How a document-processing outsourcer replaced manual keying with a configurable capture platform - per-client extraction schemas, Mistral OCR and AI pulling structured fields with a confidence rating on each, and a three-pane QC desk where an operator reviews the extraction against the original document before anything is exported.
The situation
Document-processing outsourcing is volume work. Invoices and similar documents arrive as PDFs, scans, JPGs, PNGs and TIFFs, and somebody has to turn them into structured data the client can use. Traditionally that means operators keying by hand for most of a working day - slow, expensive, and error-prone in a way that scales with the volume rather than against it.
The complications sit around the edges. Every client wants different fields, so an invoice project and an insurance project need different schemas. Quality control happens away from the original file, so a reviewer is checking numbers against nothing. Nothing signals which extractions are doubtful and which are safe, so every field gets the same attention. And the deliverable is Excel, which means the extraction has to be structured and auditable rather than merely readable.
Without a confidence signal, every extracted field gets the same attention - including the ones that never needed checking.
What we built
Shine is a browser-based operations platform built around one idea: let the AI read the document, and keep the person who signs it off. A project manager sets a project up through a five-step wizard, defining the extraction schema for that client - text, number, date, email, currency, long text, or table fields, each carrying its own AI instruction and ordered by drag and drop. Documents then arrive by upload, or by watched folder and FTP.
From there a background queue does the work: text pulled directly where the PDF is digital, Mistral OCR where it is a scan or an image, then a structured extraction against that project’s own schema rather than a fixed template. Every field comes back with a confidence rating, and that is what makes the review step cheap - a reviewer’s attention goes to the low-confidence fields instead of spreading evenly across the form. The QC desk itself is three panes: the document queue, a dark viewer with zoom, pan, rotate and OCR text, and the extraction form beside it. It is keyboard-first and auto-saves corrections, and it was designed to be usable across a six to eight hour shift - a different problem from a screen somebody opens twice a day. Approved records export to Excel, with table fields written as real cells.
Inside the platform
A configurable pipeline, and a review desk built for all-day use.
Capture and extraction
- Five-step wizard defining a per-client extraction schema
- Text, number, date, currency, long-text and table fields
- A separate AI instruction against each field
- Intake by upload, or by watched folder and FTP
- Direct text for digital PDFs, Mistral OCR for scans and images
- Per-field confidence, and reprocessing after an instruction is tuned
Review, export, and governance
- Three-pane QC desk: queue, document viewer, and extraction form
- Keyboard-first approval, auto-saved corrections, table cell editor
- Low-confidence fields highlighted, so attention goes where it is needed
- Assign, approve, or reject with a reason recorded
- Excel export of verified records only, table fields as structured cells
- Five server-enforced roles, audit log, and a live operations dashboard
What changed
This is a v1.0 handed to user acceptance testing, not a platform with a year of production behind it. What follows is what UAT proved, and nothing beyond it - the throughput and error-rate figures worth quoting will come from the client once the floor has been running on it.
The whole path was proven end to end
Intake, OCR, AI extraction, QC review and Excel export ran through as one pipeline in UAT, including a live Mistral run against sample invoices - among them a complex multi-column Indian invoice, which extracted correctly once the vendor-name instruction had been tuned.
Confidence scoring changes what QC costs
Because each field carries a high, medium or low rating, a reviewer works the doubtful ones rather than re-reading the whole form. That is the difference between QC as a second keying pass and QC as an exception check.
Schemas are configured, not coded
A new client project is a wizard rather than a release. Fields, AI instructions, intake path and team are all set up by a project manager, so taking on a new document type no longer waits on development.
How it was built
An npm workspaces monorepo, TypeScript end to end, with production hardening - Helmet, rate limits, a CORS allow-list, and a hard fail if production is ever pointed at local storage.
Other systems we have built near this one
How much of your document handling is still keying?
Invoices, forms, statements, claims - anywhere a person is reading a document and typing what it says into a system. Tell us what yours look like and we will tell you honestly which parts an AI can read, and which still need somebody to sign them off.