Every value carries the line it came from
Documents read, classified, extracted and validated, with confidence you can audit. Eight stages from a PDF nobody can query to structured data your system of record will accept — and a per-field confidence score deciding what a person still needs to look at.
25
document types classified in three tiers
12
cross-field validation rules, plus 16 across documents
8
stages, each replaceable without rewriting the rest
Five failure modes, and a stage against each
Every stage of the pipeline below exists to catch one of these before it reaches your system of record.
Manual data entry from unstructured PDFs
Analysts copy fields out of PDFs into the system of record by hand, one keystroke at a time.
OCR falls apart on real documents
Off-the-shelf OCR misses values in tables, stamped fields and multi-column layouts — exactly where the important numbers live.
Naive extraction invents fields
A single prompt against a document will produce plausible-looking dates, account IDs and amounts that were never in it. You cannot put that near a system of record.
Each document treated as an island
The same claimant, contract or patient is re-extracted from scratch every time, so cross-document context is lost and duplicate entities pile up.
Drift breaks production silently
A vendor updates a form template and extraction quality falls without an error being raised. The damage accumulates downstream for weeks.
The eight-stage pipeline
Each stage is a named component with its own evaluation set. You can swap any one of them without rewriting the rest.
Parse
The router selects the right parser per page rather than per document, because one scanned page in a native PDF is the common case.
Native text · OCR services · multimodal
Classify
What kind of document is this? The answer decides which schema and which rules apply downstream.
Fine-tuned classifier
Extract
The model pulls fields with high recall; deterministic patterns catch the formats that never vary; layout-aware extraction handles tables.
LLM · regex · typed schemas · layout models
Validate
A second pass reviews the extracted structure. Cross-field rules check sums, date ranges and lookups against your source systems.
Rules engine · model critic · cross-field checks
Score
Each field gets a calibrated confidence. Below threshold it goes to a person; above it, it proceeds. That threshold is a dial you own.
Confidence calibration · per-field thresholds
Graph
Entities — people, organisations, accounts, claims, contracts — are deduplicated and stored as nodes with their embeddings alongside.
Graph store · vector index
Link
New documents are linked into the existing graph, so an amendment finds its contract and a follow-up finds its original.
Cross-document reference resolution
Output
The final payload is shaped to your downstream contract, so the system of record receives what it expects rather than what we produced.
Structured JSON · REST · webhooks
Confidence comes from evidence, not self-assessment
Every extracted value carries its source quote and page. A value without a traceable source is not an answer; it is a guess.
The buyer's real question is what happens when it is wrong, and the answer is that it says so: per-field confidence, alternatives queued for a person, and nothing silently written to the system of record. On one engagement, scanned financial statements extracted at roughly 12% accuracy with text-only parsing; a multimodal change to the parse stage took the same documents to roughly 94%. Same model, different pipeline — which is the whole argument for building the stages rather than buying a box.
Six capabilities, wired into one pipeline
No black box. Every component is named, swappable, and runs in your own environment.
Hybrid parser stack
A router picks the parser per page: native text extraction where the PDF has text, OCR services where it does not, and multimodal handling for pages that are effectively images.
Fine-tuned classifier
A lightweight classifier trained on your document taxonomy routes each file to the right extraction schema, so a contract is not read with an invoice's expectations.
Model, regex and schema together
The model pulls fields with high recall; deterministic patterns and typed schemas enforce shape; a second model pass reviews the output as a critic rather than an author.
Confidence-calibrated validation
Cross-field consistency rules — date ranges, sum checks, lookups against your own records — plus a per-field confidence score, so low-confidence values route to a person.
Entity resolution and linking
Extracted entities are deduplicated and linked across documents, so the same claimant, matter or patient is recognised rather than recreated.
Drift gates and an eval harness
Every deployment ships with a labelled eval set and CI gates. When quality drops on a critical field, the pipeline blocks the release instead of shipping it quietly.
The same pipeline, configured per domain
Schemas, integrations and compliance posture change. The architecture does not.
Healthcare
Medical record abstraction
Discharge summaries, operative notes and lab reports parsed into standard clinical resources, with de-identification running at the parse stage before anything leaves your environment.
Insurance
First-notice-of-loss automation
Loss packets — reports, repair estimates, photos, standard forms — extracted into structured claim records and pushed into the claims system.
Legal
Contract abstraction at portfolio scale
Agreements abstracted to a structured schema and linked to the matter in your document system, with amendments and exhibits resolved to their parents.
Financial services
Investment document parsing
Statements, subscription documents and capital-call notices parsed into an accounting system, with entity resolution linking each one to the right holding.
Government and public sector
Form intake and case files
Citizen submissions, request packets and case files routed to the right downstream system with confidence-scored fields, deployable on-premises.
Every milestone has something you can grade
The evaluation harness is a deliverable, not an internal tool: your team grades the output rather than taking our word for it.
Week 1
Discovery and schema design
We sample 50 to 100 of your real documents, design a typed schema per document type, and define the eval set with you.
Week 2
Prototype
A working end-to-end pipeline on your sample corpus, with an evaluation harness your team can use to grade every field itself.
Weeks 3–6
Production deployment
The pipeline deployed into your own environment, integrated with your systems of record, with drift gates and evaluation running in CI.
Week 6+
Hand-off and monitoring
Runbooks, coverage and regular drift reports. Your team owns the schemas; we own pipeline reliability until you choose to take it in-house.
Bring us your hardest documents
This is our AI engineering practice
It is real work and it is where our four products came from. But what Cognilium leads with is narrower: optimization apps that run in tandem with Microsoft Dynamics 365, computing the decisions the ERP records but does not derive — the optimal price, the optimal pick path, the optimal stock level. See the optimization apps · How we build inside the ERP.