ML inference in credit scoring

Modernizing Credit Scoring Through ML Integration: A Play Telecom Case Study

Play, a telecom company with 13 million customers, implemented Nussknacker for their credit scoring system.

Nussknacker enabled the integration of MLflow into the credit decisioning process, allowing Python-based machine learning models to be deployed without developer involvement

Model inference was decoupled from the initial application, allowing data scientists to work in Python while storing models in MLflow.

Domain experts and business analysts could directly modify ML models within Nussknacker (used to calculate scoring and other business rules), reducing reliance on technical teams.

Ml models in production business rules

Business Context

Telecommunications companies face inherent risks when offering phones on installment plans and service contracts. These risks include:

  • Credit risk from customers unable to meet payment obligations
  • Fraud risk from individuals signing contracts with no intention to pay
  • Requirement to adjust credit scoring models based on evolving customer behavior and the increasing sophistication of fraudsters who devise new methods to bypass security measures

Initial Challenges

ML Model Integration Barriers

The original credit scoring system was a component of a fairly monolithic application, which created several operational challenges:

  • Model updates required developer intervention, leading to slow iteration cycles
  • Pre-processing and post-processing logic was tightly coupled within the system
  • Model deployment to production was complex and time-consuming
  • Integration of new models required significant development time

Limitations of Business Logic Changes

The system's structure contained frequently changing business logic which was maintained by developers

  • Domain-specific business rules were difficult for developers to understand and implement
  • Credit risk analysts couldn't directly manage or modify business logic
  • Frequent rule changes required detailed specifications and developer involvement

ML Solution Implementation

Direct ML model inference within Nussknacker

Implemented solution combines MLflow model registry with Nussknacker's ML runtime. This integration enabled seamless deployment of Python-based machine learning models without requiring developer intervention.

ml model inference business logic

Process Transformation

The new system architecture separated model inference from the core application, allowing data scientists to work in their preferred Python environment while storing models in MLflow.

Domain experts gained the ability to modify models and versions directly through Nussknacker's interface.

Business Logic Migration

Play moved frequently changing business rules into Nussknacker scenarios, enabling business analysts to manage decision algorithms directly. This included rules for customer eligibility, calculate deposit amount, or determing whether customer needs to provide additional documents to prove their identity.

Real-time ML Technical Architecture

The new system architecture includes Nussknacker with several key components:

  • SQL enrichers for accessing pre-calculated features
  • Real-time feature computation capabilities
  • Decision tables for channel-specific parameters
  • ML enrichers for model inference
  • Flexible response preparation and handling

Detailed Case Study

If you want to dive deeper into the technical side of this solution, go to the blog post

 

Results

The implementation of Nussknacker's platform delivered several key improvements:

Operational Efficiency

- Faster model deployment and updates
- Streamlined business rule modifications
- Reduced technical bottlenecks

Team Empowerment

- Data scientists can build models using familiar Python tools and gained autonomy in model development
- Business analysts could directly manage decision logic
- Reduced dependency on development teams

Technical Benefits

- Flexible MLflow integration for model registry and versioning
- Simplified architecture
- Enhanced monitoring capabilities
- Real-time feature computation and decision automation

Future Capabilities

The new system positions Play for continued innovation in credit risk assessment. The platform's flexibility allows for gradual rollout of new models, with capabilities for A/B testing and external service integration.

More resources

Blogpost

Blogpost

have questions?
talk to us