Turning paper into data.

Structure from the unstructured.

A document is an argument the author made with layout and prose. A human reader reconstructs the argument in their head — this is the party name, this is the date, this clause modifies that one, this number is a total and that number is a line item. The reconstruction happens fast and unconsciously, which is why it feels like nothing until you have to do it on 200 invoices. Then it is everything.

Document intelligence is the discipline of getting a machine to do that reconstruction at volume, with enough ground-truth checks that a human only needs to audit the uncertain cases. Vision models do the reading. Structured-output schemas enforce the shape of the answer. Confidence scores tell you what to trust and what to double-check. The output is not a summary — it is a dataset.

Model, schema, and a reviewer.

The extraction stack has three layers that have to work together. The vision-capable model does the actual reading — looking at the pixels, resolving the layout, parsing the prose. The structured-output schema enforces the shape of the answer, so the model can't hallucinate new fields or omit required ones. And a human reviewer audits the confidence-flagged cases before they enter the system of record. Skip any one layer and the output becomes unreliable; skip two and the whole pipeline is worse than doing the work by hand.

From inbox to database row.

A good pipeline is a conveyor belt, not a crane lift. Each stage takes a standardised input and produces a standardised output, so the next stage can start without knowing what happened before. The whole thing runs on autopilot 95% of the time; the human only gets involved at the exceptions.

The jobs that actually earn the stack.

These are the extraction patterns that come up again and again in an industrial service business. Each one has its own schema, its own validation rules, and its own failure modes. Together they cover most of what the office does with paper on any given week.

Drop in a document, watch it come apart.

This demo uses pre-computed synthetic extractions so the page works offline and preserves privacy — no API keys, no backend, no uploads sent anywhere. The structured outputs below are what a real Claude-powered pipeline produces on similar documents; the formatting, field shape, confidence scores, and red-flag logic all mirror production behaviour.

"Page 6 of 23 said everything."

A standard supplier purchase agreement arrives in the office queue. Twenty-three pages, three signatories, English throughout, nothing visually alarming. The procurement lead runs it through extraction because that's the routine now — she's not expecting anything. The output comes back with a single amber flag: lead time, 14 weeks, confidence 71%, flagged for review.

Why this is the AI leverage. A procurement lead reading 23 pages at the end of a long day skims the commercial terms block — the important stuff is supposed to be on page 1, everyone knows that. Extraction reads every page with equal attention and every paragraph with equal scepticism. It doesn't skim. It doesn't get tired. And when the phrasing is unusual, it flags the uncertainty rather than guessing confidently — which is exactly what a careful human reader would do if they had eight hours for every contract instead of eight minutes.