The situation
The agency reviews roughly 700 independent public audits per year, against decades of accumulated filings. Every review was 100% manual, consuming 4 to 24 hours per audit due to non-standardized layouts, siloed data systems, and manual balance sheet math checks. The hardest part was not this year's submissions but the back-file: years of legacy audits in formats that had changed repeatedly over time, scanned at varying quality, with no consistent structure to key on. A reviewer could not compare a filing against its own history without opening the history by hand. Previous commercial AI tools failed with high false positive rates in budget schedules.
What we built
In an 8-week compressed build under a strict fiscal-year deadline, we built 'Reviewer Zero'. We established a 3-layer architecture: Layer A (deterministic bytes classification & financial facts ledger), Layer B (reasoning agent for compliance guidelines), and Layer C (faithfulness gates & 7-year audit retention locks).
- 100% deterministic financial ledger math checks (zero model hallucination)
- Dual-citation verification: resolves audit text and governing standard
- Content-based word geometry highlight resolution instead of page numbers
- Product contract: AI never authors letter comments; humans write feedback
- 7-year immutable compliance retention lock on all generated reports
- Live API integration with third-party audit submission portal
Stack shipped
Deterministic Financial Ledger · Byte-level Document Classifier · Two-Sided Citation Resolver · Faithfulness Gate (≥0.95) · Government Cloud IaC · Legacy audit archive
