Problem Statement
The South African fintech ecosystem has grown exponentially, processing millions of transactions daily. However, this growth has attracted sophisticated fraud vectors, including SIM swaps, synthetic identities, and account takeovers. Traditional rule-based systems are insufficient.
FINTECH SENTINEL addresses this by deploying machine learning models that analyze transaction patterns, user behavior, and network relationships in real-time. The system identifies anomalies that human analysts might miss and provides a secure, intuitive dashboard for investigation and response.
Technical Architecture
Data Ingestion Layer
Transactions flow from partner banks into Apache Kafka topics for high-throughput ingestion. Validated and normalized using Apache Spark Streaming.
AI/ML Processing Layer
Parallel models: Isolation Forest detects outliers, XGBoost predicts fraud probability, and a Graph Neural Network analyzes transaction networks for coordinated fraud rings. Auto-retrained on Vertex AI.
Frontend Dashboard Layer
React/Next.js application displaying real-time alerts, scores, and network visualisations. WebSockets ensure live updates without refreshes.
Technical Stack
AI / Machine Learning
Python 3.11, TensorFlow 2.x, Scikit-learn, XGBoost, PyTorch (GNN), Kafka.
Frontend Architecture
React 18, Next.js 14, TypeScript, D3.js (Network Vis), Material-UI, Tailwind CSS.
Backend & Data
FastAPI, PostgreSQL, Redis, JWT Authentication.
Cloud Infrastructure
GCP Vertex AI, Dataflow, Pub/Sub, BigQuery, Cloud Run.
Impact & Metrics
73%False Positives Reduced
89%Fraud Detected
3mAvg Investigation Time
Code Architecture: ML Pipeline
# fraud_detection/model_pipeline.py
import pandas as pd
import numpy as np
from sklearn.ensemble import IsolationForest
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense, Dropout
class FraudDetectionPipeline:
def __init__(self):
self.isolation_forest = IsolationForest(contamination=0.05)
self.lstm_model = None
def predict_fraud_score(self, transaction: dict) -> float:
""" Generate fraud score (0-1) combining multiple models """
if_score = -self.isolation_forest.score_samples([transaction['features']])[0]
if_score = (if_score - if_score.min()) / (if_score.max() - if_score.min())
lstm_score = self.lstm_model.predict(transaction['sequence'])[0][0]
fraud_score = 0.4 * if_score + 0.6 * lstm_score
return float(fraud_score)
