SentinelGraph: Real-Time Hybrid Fraud Detection
Welcome to the Live Anti-Fraud Lab.
This interactive simulator stress-tests an MLOps pipeline designed to intercept fraudulent transactions at scale. SentinelGraph deploys a hybrid architecture that evaluates transactions through three concurrent paradigms in real time:
- Deterministic Layer: Static heuristic rules for instant edge filtering.
- Gradient Boosting Layer: A XGBoost classifier engineered for tabular feature patterns.
- Deep Learning Layer: An unsupervised Autoencoder neural network dedicated to structural anomaly detection.
This is an ongoing story, want to know more?
READ THE FULL ARTICLE →Live Simulator
Transactions analyzed
0
Losses avoided
€ 0.00
Losses incurred
€ 0.00
Vanity accuracy
100.0%
Enterprise snapshot
Reusable backend analytics for both FastHTML and the external Astro frontend.
Total predictions
0
Persisted decisions in PostgreSQL
Alerts today
0
ALERT + BLOCK from the current UTC day
Blocked total
0
Historical hard blocks across all stored runs
Database status
Unavailable
Shared source of truth for dashboards and blog widgets
Recent decisions
Operational trail from `/api/analytics/recent`.
No persisted predictions yet.
Recent simulation runs
Stored scenario batches help separate live serving from evaluation logic.
No simulation runs recorded yet.
XGBoost evaluation
Offline holdout metrics explain model quality; operational analytics explain live behavior.
Precision
0.990
False positives under control
Recall
1.000
Coverage of actual fraud cases
F1-score
0.995
Balanced operating quality
ROC AUC
1.000
Ranking quality across thresholds
PR AUC
1.000
More honest under class imbalance
Offline holdout evaluation for XGBoost. Use alongside operational analytics from live simulations.
Confusion matrix
True negatives
56863
Legit transactions correctly passed
False positives
1
Customer friction from unnecessary alerts
False negatives
0
Missed frauds with real business cost
True positives
98
Frauds correctly captured
Threshold trade-off
Recommended threshold 0.95 maximizes F1 at 0.995.
FP rate 0.000 · FN rate 0.000
Threshold vs F1
ROC curve
Precision-Recall curve
SHAP explainability
Mean absolute SHAP impact for top XGBoost features on the offline holdout sample.
Business interpretation
False negatives miss fraud and burn money. False positives create friction. Recall and F1 therefore matter more than raw accuracy in Sentinel.
Holdout rows: 56962 · positive rate: 0.0017
Operational widgets above explain what the engine is doing now; these offline metrics explain whether the XGBoost wall is actually trustworthy.
Live transaction stream
HTMX polling across 3 walls — Rules, XGBoost, Anomaly Detection.
Request sent — waking up the anti-fraud engine…
Transaction feed
Waiting for the first tick…
🛡️ The 3 walls are analyzing the transaction…
Latest transaction details will appear here after each tick.