How the technology works
Cognaize is neuro-symbolic AI built for finance. Neural models read your documents. Symbolic models verify every output against your institution's rules, and the system keeps working until every check passes.
Two kinds of intelligence, one system
Cognaize merges two kinds of AI: neural models that adapt to messy documents, and symbolic models that enforce your rules. A workflow engine coordinates them to keep the system efficient at scale.
Models that learn from examples. Language models (LLMs and SLMs) read the text; layout and vision models read tables, structure, and charts.
Hard rules and accounting identities, written as logic. Every check either passes or fails, and the same input always gives the same answer.
Adaptive where documents are messy. Exact where the rules demand it.
Why not use an LLM to analyze financial documents?
LLMs reason by similarity, not rules. They place "cash" near "restricted cash" because the words appear in similar contexts. Useful for reading language, unreliable for financial data.
LLMs are non-deterministic. Finance requires deterministic outcomes.
Running large models on every page of every document gets costly fast at institutional volume.
Your ontologies, embedded in the AI
Every institution interprets data differently. Cognaize captures your schema, terminology, and rules, then enforces it on every document.
Define your ontology once
Your schema, terminology, and rules, written in plain business terms.
The Semantic Kernel translates it
Plain language becomes formulas in code the system can run.
Every document validated
Applied automatically, on every page. Same input, same output: every check on record.
From document to data
-
Read
Finance-tuned models (OCR, layout, vision, SLMs) extract text, structure, and visuals from any document quality.
-
Structure
The system creates a version where tables, footnotes, and text become connected elements.
-
Contextualize
Field-by-field extraction adjusted to your schema and definitions.
-
Verify
Every value is checked against your rules. Any failure triggers a re-run or analyst review.
-
Orchestrate
Feed usable data, in any format, to where your systems live (e.g. JSON, Excel, CSV, etc.).
The cheapest path that passes
The Hypergraph Workflow Engine starts every document on the lightest models and escalates only when a check fails.
Most documents never need the heavier models.
Intermediate results (OCR text, detected tables) are saved and reused.
Models built for finance
Cognaize models are trained on over 8 million fully annotated financial documents, each labeled by CFAs.