Use case · Scanned-page tagging
Tagging scanned pages at the pace they arrive.
A token factory for the less glamorous half of a real engineering archive: scans. OCR plus classification plus taxonomy stewardship, under one owned architecture.
The problem
What we are actually solving.
A working archive is never tidy. Alongside the structured technical documentation, aircraft operators accumulate scanned material — historical records, handwritten work cards, paper trails from suppliers. This content does not arrive with a schema; it arrives as pixels.
The operational question is simple and relentless: what is on this page, and what category does it belong to. Answering it manually does not scale. Answering it with a naive OCR-plus-chatbot pipeline produces confident nonsense, which is worse than manual handling, because nonsense is cheap to produce and expensive to catch.
The approach
The shape of the token factory for this use case.
The context engine here is a visual parser: the unit of retrieval is a page image plus its OCR layer plus any structured metadata the source system provides. Embeddings are computed over the combined representation, not the OCR text alone, so degraded scans do not silently dominate the distribution.
Classification is a multi-stage pipeline inside the factory. A triage step routes pages by broad type; a second stage applies the fine-grained taxonomy; a final stage emits the structured fields the downstream systems expect. Each stage is an owned prompt, a routed model call, and a gated eval — the same four-layer shape as the text pipeline, adapted to visual inputs.
Taxonomy stewardship is part of the engagement, not a side effect. The classification labels are a living schema, owned by the operator, versioned in the factory, and replayed against the eval set on every change.
Outcomes
What changed.
Pages tagged
Hundreds of thousands
Across historical and incoming material
Classification accuracy
High-nineties
Top-label, against human adjudication
Manual-tagging time
Largely eliminated
On the routine share of incoming volume
The operational win was steadiness: new scans stopped piling up in a queue and started flowing through the pipeline at the rate they arrived. Human reviewers shifted from data entry to exception handling — which is where the judgement the organisation was paying for actually lives.
The taxonomy became an asset in its own right. Once the labels were owned, versioned, and measurable, other systems in the organisation started reading them. That was not in the original brief.
Honesty
Delivered engagement. Same company and multi-year horizon as the technical documentation work. Happy to walk through specifics — taxonomy shape, error-bar handling, the particular pathologies of scanned material — under NDA.
Rounded metrics are used here because formal numbers were not externally published. The shape of the result is stable; the decimal places are not the point.
Next step
Talk about tagging scanned pages at the pace they arrive..
The usual first engagement is an architecture review. Say so on the form, or write directly.
Get in touch