Catch Me If You Can
A 5-level intelligent cascade that routes fraud decisions to the cheapest sufficient processing layer. Rules when possible. AI when necessary.
Problem
Rules are fast but rigid. They catch known patterns in milliseconds but miss sophisticated fraud.
AI is smart but expensive. At $0.03–0.05 per decision, pure AI costs $30–50K monthly for 1M transactions.
The cascade routes each transaction to the cheapest layer that can handle it - only escalating when the expected ROI justifies it.
Architecture
| Layer | Latency · Cost · Volume | |
|---|---|---|
| 01 | Rules Engine | <50ms · $0.0001 · 70% |
| 02 | Statistical ML | 1–2s · $0.001 · 20% |
| 03 | Single AI Agent | 2–3s · $0.01 · 7% |
| 04 | Prosecution on Demand | 3–5s · $0.03 · 2% |
| 05 | Adversarial Courtroom | 5–10s · $0.05 · 1% |
Key Ideas
ROI-based routing. Escalation decisions are economic. A $50K wire at 75% confidence has $12,500 expected loss - worth spending $0.03 on AI. A $5 coffee isn't.
Self-learning rules. When expensive layers catch fraud that cheap layers missed, the system extracts the pattern and proposes a new rule. Over time, processing migrates from $0.05 to $0.0001 per transaction.
Domain agnostic. One architecture for finance, healthcare, insurance, cybersecurity, and supply chain. Domain-specific knowledge loads dynamically - the cascade logic stays the same.
Evolutionary adversarial testing. A genetic algorithm generates attack patterns to probe defenses. Population of 20 patterns, elite selection, 70% crossover, 30% mutation - amount jittering, merchant variation, time shifting. The strongest attacks train the cascade. The system evolves against itself.
Technical Disclosures
Defensive publications via TD Commons.
Multi-Stage Failure Learning (APLS)
Automated rule generation from cascade failures. When upper layers catch fraud that lower layers missed, the system extracts patterns, generates deterministic rules, and migrates processing to cheaper tiers.
ROI-Based Cascade Routing
Predictive cost-benefit analysis for escalation decisions. The system calculates expected fraud loss versus processing cost, routing only when ROI exceeds 50×.
Adaptive Multi-Agent Debate Framework
A tiered reasoning framework for debate-driven decision-making. Multiple AI agents generate and evaluate arguments with structured argumentation and dynamic evaluation criteria - powering Levels 4-5 of the cascade.