The model recommends. Deterministic code decides. A human signs off.
An AI agent doesn't get to move money, approve an invoice, or overwrite a governing contract clause just because it sounds confident. Every system below follows the same architecture: a Primary Agent extracts and recommends, a Critic Agent validates that recommendation against deterministic rules, and nothing reaches production state until a human signs off — with every step written to a log nobody, including me, can quietly edit after the fact.
flowchart LR
A[Input<br/>Invoice / Filing / Contract / Schedule] --> B(Primary Agent<br/>Extracts & Recommends)
B --> C{Critic Agent<br/>Validates Against Rules}
C -->|Fails validation| B
C -->|Passes validation| D[Deterministic Decision Layer]
D --> E([Human Sign-Off])
E --> F[(Immutable / Hash-Chained Audit Log)]
style B fill:#2d5cf7,color:#fff
style C fill:#c74e33,color:#fff
style D fill:#1a1a1a,color:#fff
style E fill:#2f5233,color:#fff
This isn't one project's diagram — it's the constraint every project below is built inside.
| Project | What it governs | Verified result |
|---|---|---|
| ReconcileAI | Month-end close reconciliation | 97.5% automated match rate · 40+ hrs saved/cycle · 100% audit coverage |
| AuditLedger | AP 3-way match & chargeback | 100% error detection, 0 false auto-approvals · 60% straight-through |
| StratBrief AI | SEC filing → board memo | Zero numeric drift · full paragraph-level citation across 5 tickers |
| ClauseKinetic | Contract precedence & chargeback exception | Amendment-precedence engine · 2 production security gaps closed pre-launch · adversarially chaos-tested before shipping |
| ShiftProof AI (Rostrix) | Staffing & scheduling resilience | 528,000-run Monte Carlo stress test · self-audits and labels every output MEASURED, ESTIMATED, or ASSUMPTION |
NIST AI RMF 1.0 · Bloomberg Market Concepts · IAPP AIGP (in progress) · Lean Six Sigma Green Belt (in progress)
Python LangChain / RAG SQL Streamlit Azure AI PostgreSQL FastAPI