ML ENGINEERING · 2026

AI Fraud Surveillance

A real-time market surveillance platform that flags wash trading, spoofing and quote stuffing with a three-model AI ensemble on an Azure Delta lakehouse, with live dashboards for analysts.

AI Fraud Surveillance architecture diagram
Real-time market surveillance engine: ingestion, Databricks medallion lakehouse, multi-model AI ensemble and the surveillance action layer.

Overview

Algorithmic market abuse such as wash trading, layered spoofing and quote stuffing happens in fractions of a second, while traditional compliance checks run as end-of-day batch audits. This platform closes that gap with real-time surveillance.

Market telemetry streams into an Azure Databricks Delta lakehouse, 43 microstructural features are computed live, and a three-model AI ensemble scores every order. Each flag is explained with TreeSHAP so analysts can see exactly why it was raised.

What it does

01

Three-model ensemble

Supervised XGBoost for known attack patterns, Isolation Forest for statistical anomalies and a deep autoencoder for unseen patterns.

02

Explainable flags

Every flagged order comes with TreeSHAP feature attributions such as cancel ratio, cancellation velocity and order-to-trade ratio.

03

Medallion lakehouse

Bronze raw ingestion, Silver cleaning and deduplication, and a Gold governed feature store on Delta Lake.

04

3D surveillance console

A WebGL radar console with a live audit ledger, verdict filters and instant threat alerts.

05

Crypto market terminal

Order book depth, buy/sell pressure imbalance and cross-market correlation to spot coordinated pump-and-dump activity.

06

Executive dashboards

Power BI on the Gold layer with KPI measures for enforcement rates, false positives and SLA compliance.

How it works

  1. Ingest

    Order flow, order books and trade prints stream through Azure Event Hubs into the Bronze Delta tables.

  2. Clean & engineer

    Spark notebooks validate, deduplicate and sequence events, then compute 43 microstructural features in the Gold layer.

  3. Score

    The ensemble combines the three models into one composite risk score for every transaction.

  4. Explain & act

    TreeSHAP explains each flag, and results flow to the web consoles and Power BI for analysts.

Preview

AI Fraud Surveillance preview

Tech stack

Machine learning

XGBoostIsolation ForestPyTorch autoencoderTreeSHAPMLflow

Data & cloud

Azure Event HubsAzure DatabricksDelta LakeSparkTerraform

Interfaces

Three.js / WebGLChart.jsPower BIVercel edge functions
NEXT PROJECT →Zurbez StorefrontE-commerce Platform