3 Real ML Models · Live Pipeline · Full Explainability

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.

XGBoostLightGBMIsolationForestKeras LSTMSHAPFastAPI
Live Transaction Stream
5 / 1k shown

Amazon.com

just now

$24.99SAFE

Wire Transfer

2s ago

$4,200.00BLOCKED

Uber Eats

4s ago

$89.50SAFE

Crypto Exchange

6s ago

$2,800.00REVIEW

Netflix

9s ago

$12.99SAFE

12.4k/s

Processed

0.8%

Flagged

23

Blocked

0%+

Model 1 AUC

0%

Model 2 AUC

0%+

Model 3 AUC

0K

Model 1 Rows

0.0K

Model 2 Rows

0

Live ML Models

Detection Pipeline

How a Transaction Flows Through the System

Transaction Risk Score

TransactionGuard XGB+LGB

XGBoost + LightGBM ensemble scores the transaction itself — amount, velocity features, card metadata.

Anomaly Flag

AnomalyNet Hybrid

IsolationForest + Autoencoder checks whether this transaction looks statistically unusual — no fraud labels needed.

Behavioural Risk Score

BehaviourGuard LSTM

GBM + LSTM reads the user's last 5 transactions as a sequence — detects account takeover even when the transaction looks normal.

LOW / MEDIUM / HIGH / CRITICAL

Combined Risk Engine

Weighted average (40%/30%/30%) of all three scores → single dashboard alert with full reasoning.

Try the engine

Detection Models

3 Models, Fully Testable

View deep-dive charts
Transaction Fraud

TransactionGuard XGB+LGB

IEEE-CIS ensemble: XGBoost + LightGBM scoring real Vesta Corp transactions.

PIPELINE ROLE

Is this TRANSACTION fraudulent?

XGBoostLightGBMSHAPscikit-learnFastAPI

~0.95+

Ensemble AUC

XGB + LGB

Models

Structured / CSV data
Test Model
Anomaly Detection

AnomalyNet Hybrid

Hybrid Isolation Forest + Keras Autoencoder — zero-label unsupervised detection.

PIPELINE ROLE

Does this transaction look STATISTICALLY unusual?

IsolationForestKerasTensorFlowRobustScalerscikit-learnFastAPI

~0.97

Hybrid AUC

~0.88

IsoForest AUC

Structured / CSV data
Test Model
Account Takeover

BehaviourGuard LSTM

GBM + LSTM sequence model — catches account takeover by detecting USER behaviour shifts.

PIPELINE ROLE

Is this USER behaving like THEMSELVES?

GradientBoostingKeras LSTMTensorFlowRobustScalerscikit-learnFastAPI

~0.91+

Ensemble AUC

~0.87

GBM AUC

Structured / CSV data
Test Model

Under The Hood

How It Works

01

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.

02

Three Complementary Lenses

Supervised catches known fraud. Unsupervised catches anomalies. LSTM sequence model catches account takeover. Each covers blind spots of the other two.

03

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.

04

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.