Improving Real-Time Credit, Fraud, and Pricing Decisions with Hopsworks

Clicklease is a U.S.-based fintech company that provides micro-leasing solutions for small business owners. By enabling fast, data-driven credit decisions for equipment financing such as trailers, kitchen equipment, and point-of-sale assets, Clicklease helps businesses access capital they would otherwise struggle to obtain.

To support real-time use cases such as fraud detection, probability-of-default scoring, and lease pricing, Clicklease used Hopsworks to modernize its feature engineering and model serving platform.

The Aim

As Clicklease scaled its machine learning workflows, several challenges emerged in its existing architecture:

• Heavy reliance on stored procedures in Postgres for feature logic
• Large third-party payloads flattened and scanned at request time, leading to high latency
• Feature logic tightly coupled with application code and model deployments
• Inconsistent feature transformations between training and production
• Training–serving skew and data contract violations
• Limited reusability of feature logic across models
• Manual, risky model updates with shared dependencies
• High operational overhead for a small engineering team

Clicklease needed a platform that could support on-demand feature computation, low-latency serving, and independent model deployment without increasing operational complexity.

Why Hopsworks?

Clicklease adopted the Hopsworks Feature Store as a centralized platform for managing features and serving models in real time.

With Hopsworks, Clicklease was able to:

• Compute and serve on-demand features at request time without expensive database scans
• Achieve predictable, sub-second latency for real-time decisioning
• Use a unified feature schema for offline training and online inference
• Decouple feature engineering from application and model deployment logic
• Centralize feature definitions and transformations in Python
• Leverage built-in schema validation to reduce runtime errors
• Manage models independently using the Hopsworks Model Registry and serving layer (built on KServe)
• Log features and predictions automatically for monitoring, auditing, and backfilling
• Reduce operational burden for a lean platform and MLOps team

Results

Faster, Reliable Real-Time Decisions

• Sub-second feature serving for credit, fraud, and pricing models
• Removal of heavy database scan queries during inference
• More predictable latency in customer-facing approval flows

Consistent Features Across Training and Serving

• Same feature transformations used offline and online
• Reduced training–serving skew and data contract issues
• More reliable production model behavior

Simplified Model Operations for a Lean Team

• Independent model updates without redeploying application services
• No shared dependencies across model deployments
• Lower operational overhead for a small engineering team

On-Demand Feature Lifecycle Management

• Centralized logic for computing features at inference time
• Reusable and testable feature transformations
• Faster iteration on fraud, risk, and pricing models