01Challenge

Covenant testing is manual and costly work for a credit analyst: find the agreement version in force, extract the formula, gather figures from financial statements and transactions, compute and justify the conclusion. A wrong number is expensive, so the AI's output has to be reproducible and verifiable.

02Solution

• A six-stage pipeline: PDF ingestion, document routing, extraction of covenants, facts and transaction categories, computation and self-audit
• Scanned pages without a text layer are read by a vision model
• The model turns each covenant into an expression over a fixed vocabulary of aggregates, and a Python engine evaluates it — all arithmetic is deterministic
• Strict JSON schemas at every stage and an on-disk cache of model responses keyed by request hash
• A full trace: document type, extracted facts, the category of every transaction, formula inputs and evidence candidates

03Results

• 97.2% match with the reference on the public set — 35 of 36 cells
• A hidden set with a different structure — 27 borrowers, 84 cells — with zero critical findings
• Self-audit blocks the submission if any cell is not justified
• Re-runs are almost free thanks to the cache