Risk Scoring Agent — Real-time Transaction Risk Assessment
ML-powered risk: feature extraction → model inference → score fusion → decision engine → continuous learning · <50ms latency
1 Feature extraction & enrichment
Transaction features
Amount (normalized)
Currency pair
Transaction type
Time of day
Day of week
User features
Account age
Verification level
Historical volume
Avg transaction size
Risk tier
Behavioral features
Velocity (1h/24h/7d)
New recipient flag
New device flag
New corridor flag
Session anomalies
Network features
IP geolocation
IP reputation score
Device fingerprint
VPN/proxy detection
Browser signals
External enrichment
Watchlist screening
PEP check
Adverse media
Sanctions lists
Consortium data
Feature extraction: <5ms
2 ML model inference
Fraud model
XGBoost ensemble
Transaction fraud prob
Account takeover prob
Synthetic ID prob
First-party fraud prob
AML model
Graph neural network
Structuring score
Layering score
Integration score
Network risk
Anomaly model
Isolation forest
Behavioral anomaly
Amount anomaly
Pattern deviation
Cluster distance
Credit risk model
Gradient boosting
Default probability
Exposure at default
Loss given default
Expected loss
Model serving
TensorFlow Serving
Model registry
A/B testing
Shadow mode
Fallback rules
Inference: <20ms
3 Score fusion & calibration
Score combination
Weighted ensemble
Fraud: 40% weight
AML: 30% weight
Anomaly: 20% weight
Credit: 10% weight
Risk score output
Low (0-30)
Auto-approve
Medium (31-70)
Enhanced
High (71-100)
Review
Confidence interval
Explainability
Calibration
Platt scaling
Isotonic regression
Threshold tuning
Segment calibration
Daily drift check
Explainability
SHAP values
Top risk factors
Rule triggers
Analyst summary
Audit documentation
Explainability: always
4 Decision engine & actions
Decision rules
Rule engine
Blacklist check
Velocity limits
Amount limits
Corridor rules
Decision outcomes
APPROVE
STEP-UP
REVIEW
DECLINE
BLOCK
Step-up actions
2FA challenge
OTP verification
Selfie verification
Document request
Phone callback
Review queue
Priority scoring
Analyst assignment
SLA tracking
Decision tools
Case management
Regulatory actions
SAR filing trigger
CTR filing
Account freeze
LE notification
Exit procedure
Decision: <50ms E2E
5 Continuous learning & monitoring
Feedback loop
Outcome labeling
Analyst decisions
Chargeback data
SAR confirmations
False positive tags
Model retraining
Weekly retrain
A/B champion
Validation set
Bias testing
Canary deploy
Performance metrics
Precision/recall
AUC-ROC
False positive rate
$ saved
Review reduction
FPR target: <2%
Monitoring
Score distribution
Feature drift
Concept drift
Latency tracking
Alert thresholds