Read, extracted, and checkable
Contracts, financial documents, forms and scans turned into structured data a business can act on. With confidence you can audit, because every value is tied to the exact place in the document it came from.
25
document types classified, in three tiers
12
cross-field rules, plus 16 across documents
0
values stored without a traceable source
Twelve steps, eight of them shown
Built for a family-office platform and generalised since. The validation stages are the ones that decide whether the output can go near a system of record.
Upload and parse-quality detection
PDF, DOCX, spreadsheet or scan — and the first decision is what kind of parsing this file actually needs: native text, mixed, or full OCR.
Classification across 25 types
In three tiers, because a contract and an invoice should not be read with the same expectations, and neither should two kinds of contract.
Schema-driven extraction
A registry holds a schema per document type, and extraction is driven by it rather than by a prompt that hopes for the right shape.
Field alias mapping
The same field is called four things across four issuers. That mapping is data, not a special case in code.
Cross-field validation
Twelve rules checking that the values agree with each other — sums, date ranges, and lookups that should reconcile.
Cross-document validation
Sixteen more rules and six edge types linking documents to each other, so an amendment finds its contract.
Evidence-based confidence
High, medium or low, with the source quote attached — and up to three alternative readings per field where the document is genuinely ambiguous.
Graph and document storage
Entities resolved and stored so the next document about the same party is recognised rather than recreated.
What confidence actually means here
Confidence comes from verifying the evidence exists, not from the model grading itself.
Every extracted value carries its source quote and page. A value without a traceable source is not an answer; it is a guess. And the buyer's real question — what happens when it is wrong — is answered by the design: it says so, per field, with alternatives queued for a person. 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.
Contract intelligence, for legal teams
We build systems that change how legal teams review, negotiate and manage contracts — cutting the reading, never the judgement, while making the standard applied consistent from the first contract to the fortieth.
Read against your own playbook
A firm's legal playbook is parsed into 80 to 150 enriched terms, and every clause is checked against those positions rather than against general legal knowledge.
Findings, not raw text
Risk flagged per clause with replacement language already drafted, so a lawyer reviews findings instead of reading pages to locate them.
Cited to source, always
Every finding points at the clause that produced it. A conclusion a partner cannot trace is a conclusion they will not sign.
This is not a proposal: it is Paralegent AI, our own product, in production today — 23 agents reviewing a contract inside Word in 5 to 10 minutes. See Paralegent AI.
The same pipeline, different documents
Including one place where determinism beat the model: an Excel playbook is parsed with zero model calls, because the structure was already there.
Contract review
A company's legal playbook parsed into 80 to 150 enriched terms, then applied clause by clause inside Word.
See it →The full eight-stage pipeline
Parse, classify, extract, validate, score, graph, link, output — each stage replaceable without rewriting the rest.
See it →Retrieval over what you extracted
Once documents are structured, the questions people actually ask become answerable.
See it →Send us the documents nobody wants to key in
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.