The Chaos of Financial Formats
Commercial banks rely on standardized financial "spreading" to feed credit risk models. However, the source documents vary wildly—from structured Canadian T2 tax returns to informal, handwritten micro-ledgers in West Africa.
We successfully deployed the Cognitive Engine across three distinct financial environments:
- DUCA Financial (Canada): Bilingual English/French T2 returns and Notices of Assessment.
- Alrahma (Micro-Credit): Informal, non-standardized ledgers from unbanked entrepreneurs.
- BSIC (Benin): OHADA-compliant records across diverse regional accounting standards.
Cognition for Reasoning. Math for Math.
Our breakthrough was separating the workflow: Generative AI for reasoning and reclassification, and Deterministic engines for calculation. This ensures 100% accuracy in financial ratios.
-
Lossless Ingestion
The Engine captures the statement exactly as presented. If a client provides 40 unique line items, we ingest all 40. No initial assumptions. -
Cognitive Reclassification
Specialized agents perform expert-level mapping via the Foundry. They combine equivalent items (e.g., Cash + ST Reserves) based on specific bank policy. -
Cross-Agent Validation
Multiple agents perform the tasks independently and cross-check outcomes, flagging only the low-confidence outliers for human review. -
The Deterministic Math Guard
We never ask an LLM to do math. Once data is standardized, a traditional calculation engine computes Leverage, EBITDA, and Liquidity ratios. -
Audit-Ready Spreading
Every reclassified dollar is linked back to the original pixel in the source PDF for regulatory auditability.
The "Glass Box" Edge
This is not simple OCR. This is the industrialization of financial expertise. We captured how senior analysts think about risk reclassification and Formalized it into a machine-speed workflow.