SecureNet — Real-Time Threat Detection

Problem

Financial services clients face sophisticated, fast-moving fraud. Traditional rules-based systems struggled to keep up and generated many false positives.

Approach

We combined streaming feature extraction with unsupervised anomaly detection and a lightweight supervised model that learns from analyst feedback in production.

Solution

• Streaming feature pipeline (Kafka) and low-latency feature store.
• Hybrid anomaly detection with explainability for analysts.
• Feedback loop to incorporate human analyst labels into model updates.

Outcome

• Detected 97% of fraud attempts in the evaluation dataset.
• Reduced false positives by 70%, saving analyst time.
• Production latency under 150ms for alerting.

Tech Stack

Kafka, Flink, Python, PyTorch, Redis, Docker, Prometheus for monitoring.