Patent Published

V.E.C.T.O.R

Velocity-Enhanced Clustering for Transactional Outlier Recognition

Real-Time Fraud Detection | Behavioral AI | Streaming Systems

11 Technologies·8 Key Features·2025·high complexity

Overview

V.E.C.T.O.R is a real-time fraud detection engine that personalizes anomaly detection to each user's behavior profile. Instead of applying a single global fraud model to all users, it continuously learns individual behavior, clusters users into behavioral personas, assigns persona-specific anomaly models, and produces fraud risk scores in under 50 ms. The system handles cold-start users via an XGBoost fallback, manages ambiguous transactions through 2FA verification, and provides a live monitoring dashboard. It is designed as a production-grade multi-layer streaming pipeline.

Key Features

Real-Time Transaction Monitoring

Stream-based scoring <50 ms per transaction.

Dynamic User Profiling

Continuously updated behavior profiles: velocity, average spend, frequency, session behavior, time gaps, loan activity, channel usage.

Behavioral Clustering (Persona Engine)

Users grouped into clusters (low-spend/high-frequency, high-spend/low-frequency, loan-heavy, abnormal velocity). Fraud baselines differ per cluster.

Persona-Specific Fraud Models

Each cluster has its own Isolation Forest trained on that cluster's transaction distribution.

Cold-Start Handling

XGBoost fallback for users with <10 transactions.

Decision and Verification

Fraud score thresholds: <0.4 Approve, 0.4–0.8 2FA, >0.8 Block/Manual review.

Suspicion Buffer Mechanism

Logs suspicious but verified transactions to build trust-vs-suspicion history.

Live Dashboard

React frontend with real-time charts, Cloudflare Tunnel hosting.

Technology Stack

Algorithms & Methods

UMAP

Dimensionality reduction for high-dimensional behavior vectors before clustering.

HDBSCAN

Density-based clustering, auto-detects number of clusters, supports noise/outliers.

Isolation Forest

Per-cluster anomaly detection; unsupervised, no fraud labels required.

XGBoost

Supervised fallback for early-stage scoring when user history is insufficient.

System Architecture

Producer Module

CSV producer (batch) + custom UI producer (live) → JSON transaction events

Data Ingestion (Redis Streams)

High-velocity event ingestion, consumer groups

Stream Consumer / Processing

Pull events, validate schema, compute features

Dynamic Feature Engineering

Velocity, avg amount, transactions/day, high-value frequency, time gaps, loan activity, channel usage

Clustering Pipeline (UMAP+HDBSCAN)

Build user behavior vector → UMAP reduction → HDBSCAN clustering → cluster_id

Model Management

Select model by user history (XGBoost or Isolation Forest)

Fraud Scoring

Score 0–1 from selected model

Decision Engine

Threshold logic → APPROVE / VERIFY / BLOCK

2FA Verification

OTP/email for suspicious-but-not-blocked transactions

Persistence + Audit (MongoDB)

legit_transactions and fraud_transactions collections, full audit trail

Monitoring + Dashboard

React frontend, Node.js backend, ~1-second auto-refresh, Cloudflare Tunnel hosting

Results & Outcomes

  • Real-time fraud scoring pipeline operational at <50 ms latency
  • Behavioral clustering with persona-specific anomaly detection
  • Cold-start handling via XGBoost fallback
  • Live monitoring dashboard with Cloudflare-hosted HTTPS endpoint
  • Full MongoDB audit trail for compliance

Patent

Velocity-Enhanced Clustering for Transactional Outlier Recognition

Published: 2025-10-03

Role: Primary Inventor

A patented real-time fraud detection engine that personalizes anomaly detection to each user's behavior profile using dynamic behavioral clustering, persona-specific anomaly models, and multi-layer streaming architecture with sub-50 ms scoring latency.

Related Projects