Live
XGBoost + LangChain
Total Transactions
24,891
-4.3% from yesterday
Flagged
47
+12.1% from yesterday
Blocked
12
+2 from yesterday
Model AUC
0.9847
5-fold cross-validated
Fraud Rate
24-hour rolling fraud detection rate
24h
SHAP Feature Importance
XGBoost model — top contributing features
TreeExplainer
Live Transactions
Streaming feed — updates every 2.5 seconds
live
StatusTransactionMerchantLocationRiskAmount
System Risk
Click to simulate risk change
28
Risk
18.4K
Processed
23ms
Avg Latency
93.2%
Precision
81.6%
Recall
Geographic Distribution
Transaction volume by region
Alerts
System notifications and model events
StatusTransactionMerchantLocationTimeRiskAmount
Production
XGBoost v2.1
Gradient boosted trees with scale_pos_weight for class imbalance. SHAP explainability enabled.
ROC-AUC0.9847
Avg Precision0.8623
Precision93.2%
Recall81.6%
F1-Score0.870
Latency (p95)23ms
Candidate
LightGBM v1.0
Faster training with leaf-wise growth. Under evaluation for production deployment.
ROC-AUC0.9821
Avg Precision0.8487
Precision91.7%
Recall83.2%
F1-Score0.873
Latency (p95)18ms
Baseline
Logistic Regression
Linear baseline model. Used for performance benchmarking and sanity checks.
ROC-AUC0.9712
Avg Precision0.7234
Precision88.4%
Recall72.1%
F1-Score0.794
Latency (p95)5ms
ROC-AUC Comparison
Model performance over validation folds
LLM Explanation Pipeline
LangChain integration for fraud analysis
1 Transaction features extracted from XGBoost prediction
2 SHAP TreeExplainer computes per-feature attribution
3 Top SHAP values + context sent to LangChain prompt
4 GPT-4o-mini generates compliance-ready explanation
5 Result: risk assessment + action recommendation
Model Configuration
Adjust detection model parameters and thresholds.
Risk Threshold
Transactions above this score are flagged for review
Auto-Block Threshold
Transactions above this score are automatically blocked
SHAP Explanations
Compute per-transaction feature attribution via TreeExplainer
LLM Explanations
Generate natural language fraud analysis via LangChain
Alerting
Notification rules and escalation policies.
Velocity Alerts
Alert when multiple transactions occur within a short window
Velocity Window
Time window for velocity-based fraud detection
Geo Anomaly Detection
Flag transactions from unusual geographic locations
Email Notifications
Send email alerts for critical fraud events
Data Pipeline
Transaction stream and model retraining configuration.
Stream Refresh Rate
How often the live dashboard updates
Auto Retrain
Automatically retrain model on new labeled data
Retrain Schedule
How often the model is retrained with new data