Intelligence
Against Financial
Crime.
A 3-model fraud detection pipeline — supervised transaction scoring, unsupervised anomaly detection, and LSTM behavioural sequence analysis — combined into a single risk verdict.
Amazon.com
just now
Wire Transfer
2s ago
Uber Eats
4s ago
Crypto Exchange
6s ago
Netflix
9s ago
12.4k/s
Processed
0.8%
Flagged
23
Blocked
Model 1 AUC
Model 2 AUC
Model 3 AUC
Model 1 Rows
Model 2 Rows
Live ML Models
Detection Pipeline
How a Transaction Flows Through the System
TransactionGuard XGB+LGB
XGBoost + LightGBM ensemble scores the transaction itself — amount, velocity features, card metadata.
AnomalyNet Hybrid
IsolationForest + Autoencoder checks whether this transaction looks statistically unusual — no fraud labels needed.
BehaviourGuard LSTM
GBM + LSTM reads the user's last 5 transactions as a sequence — detects account takeover even when the transaction looks normal.
Combined Risk Engine
Weighted average (40%/30%/30%) of all three scores → single dashboard alert with full reasoning.
Try the engineDetection Models
3 Models, Fully Testable
TransactionGuard XGB+LGB
IEEE-CIS ensemble: XGBoost + LightGBM scoring real Vesta Corp transactions.
PIPELINE ROLE
Is this TRANSACTION fraudulent?
~0.95+
Ensemble AUC
XGB + LGB
Models
AnomalyNet Hybrid
Hybrid Isolation Forest + Keras Autoencoder — zero-label unsupervised detection.
PIPELINE ROLE
Does this transaction look STATISTICALLY unusual?
~0.97
Hybrid AUC
~0.88
IsoForest AUC
BehaviourGuard LSTM
GBM + LSTM sequence model — catches account takeover by detecting USER behaviour shifts.
PIPELINE ROLE
Is this USER behaving like THEMSELVES?
~0.91+
Ensemble AUC
~0.87
GBM AUC
Under The Hood
How It Works
Real Training Data
Both transaction models train on real Kaggle datasets — IEEE-CIS (590K) and ULB Credit Card (284K). The ATO model engineers behavioral sequences from the same IEEE-CIS data.
Three Complementary Lenses
Supervised catches known fraud. Unsupervised catches anomalies. LSTM sequence model catches account takeover. Each covers blind spots of the other two.
FastAPI Backend
Each model is loaded from a .pkl artifact into a FastAPI server. The /predict/combined endpoint runs all three in sequence and returns individual + weighted scores.
Explainable Results
Every prediction ships with a risk gauge, verdict, reasoning text, and top contributing factors — from SHAP values, reconstruction errors, or behavioral deviation signals.
Run the Full Pipeline
Enter one transaction — all three models score it simultaneously. Watch the weighted risk engine combine their outputs into a singleLOW / MEDIUM / HIGH / CRITICALalert with per-model reasoning.