System Overview
Real-time fraud detection metrics — auto-refreshes every 30s
Live — polling /metrics & /health
—
Total Transactions
—
All time
Approved
—
—
Rejected
—
—
Under Review
—
—
Decision Distribution
Loading...
System Health
API Status—
Avg Fraud Probability—
Active Modelv1 — XGBoost
SHAP ExplainerTreeExplainer
LLM LayerGemini Flash + Fallback
Approve Thresholdprob < 0.30
Reject Thresholdprob ≥ 0.70
Cost: False Negative₹500 fraud loss
Cost: False Positive₹10 friction
Score Transaction
Submit features → XGBoost scores → SHAP attributes → LLM explains
POSTlocalhost:8000/predict
Transaction Features
Core Features
PCA Features — V1 through V28
◎
Results appear here after scoring
Transaction Explanations
Gemini-generated reasoning grounded in SHAP feature attributions — full text
GETlocalhost:8000/explanations
GETlocalhost:8000/explanation/{id}
Filter:
Loading...
Human Review Queue
HITL — ambiguous transactions (prob 0.3–0.7) routed for analyst decision
GETlocalhost:8000/review
POSTlocalhost:8000/review/{id}
Loading...
Distribution Drift Monitor
Kolmogorov-Smirnov test comparing live data against training baseline — p < 0.05 = drift
GETlocalhost:8000/drift
Loading...
Model Registry
Version history, evaluation metrics, rollback control
Registered Models
Loading...
Algorithm
XGBoost
scale_pos_weight · aucpr · 200 trees · depth 6
Explainability
SHAP + Gemini
TreeExplainer · LLM async background · fallback to rule-based
Cost Matrix
FN=₹500 / FP=₹10
50:1 asymmetry drives threshold placement