Skip to content

Decision Engine Architecture

Definition

The decision engine aggregates signals from all verification components — face liveness, face match, document authenticity, screening results — and produces a final decision: auto-approve, manual review, or auto-reject.


Decision Architecture

graph TD
    A[Verification Signals] --> B[Signal Aggregation]

    A --> A1["Face liveness score: 0.95"]
    A --> A2["Face match score: 0.78"]
    A --> A3["Document authenticity: 0.92"]
    A --> A4["Document liveness: 0.88"]
    A --> A5["Sanctions: clear"]
    A --> A6["PEP: not a PEP"]
    A --> A7["Adverse media: none"]

    B --> C[Risk Scoring Model]
    C --> D{Decision Rules}

    D -->|Score > 85 AND no screening hits| E["✅ Auto-Approve<br/>70-85% of cases"]
    D -->|Score 50-85 OR low-confidence signal| F["⚠️ Manual Review<br/>8-15% of cases"]
    D -->|Score < 50 OR screening hit| G["❌ Auto-Reject<br/>3-8% of cases"]
    D -->|Sanctions match| H["🚨 Compliance Escalation<br/><0.1%"]

    style E fill:#2E7D32,color:#fff
    style G fill:#e53935,color:#fff
    style H fill:#6A1B9A,color:#fff

Weighted Risk Score

Signal Weight Score Range
Face liveness 25% 0-100
Face match 20% 0-100
Document authenticity 15% 0-100
Document liveness 10% 0-100
OCR confidence 10% 0-100
Database verification 10% 0-100
Screening (sanctions/PEP) 10% Binary (pass/fail override)

Rules Engine

Rule Type Example
Hard rules Sanctions hit → always escalate to compliance
Threshold rules Face match < 0.50 → auto-reject
Combination rules Liveness pass + match fail → manual review
Business rules High-value account → always manual review regardless of score
Country rules FATF grey list country → require EDD

Key Takeaways

Summary

  • Decision engine combines weighted signals from all verification components
  • Hard rules (sanctions) override all scores — compliance is non-negotiable
  • Target: 70-85% auto-approve rate for efficient operations
  • Rules should be configurable per client — different risk appetites
  • Decision audit trail is legally required — every decision must be explainable