Improving a Mission-Critical Predictive Analytics Model

A specialized insurance provider relied on a custom predictive analytics model to assess claim risk, a core part of their underwriting process. As the technology landscape matured and the client’s own data grew richer, they returned to SPR for a focused four-week engagement. The goal was to make the system meaningfully better: retraining the model on stronger data, modernizing the infrastructure, and preparing the codebase for the client's next-generation internal platform.

SPR had previously built this client a machine learning model to predict claim risk across its insurance programs. The team designed it with clear documentation, sound architecture, and a focus on long-term maintainability. In production, the model was integrated into the tools underwriters used every day, helping them identify higher-risk cases and make better-informed decisions.

The model also drove measurable business outcomes. Over time, underwriters made better decisions, fewer high-risk engagements entered the pipeline, claim rates fell, and the training data began to reflect those stronger results.

Technology, especially in the age of AI, is not a one-time project. Platforms shift, data improves, and new approaches emerge. SPR’s “Evolve” phase is built for exactly this moment, to help sustain and improve AI systems so they stay aligned with changing business needs and technological advances. This engagement touched three interconnected areas:

1. Azure Infrastructure: SPR simplified the Azure architecture, removing the key vault and relocating settings that did not need to be secured to a local configuration. The result was a cleaner, easier-to-maintain system better suited to the client's emerging internal platform team, reducing ongoing operational complexity and setting up smoother future enhancements.
2. Machine Learning Model Retraining: Keeping a model effective means continuously validating it against current data. SPR retrained the model using data collected since the original build and benchmarked the results against the prior model. The retrained model showed a modest but meaningful improvement, a reflection of the original model’s real-world effectiveness. Its predictions had helped underwriters make better decisions over time, reducing high-risk engagements and lowering claim rates. Those stronger outcomes fed back into the training data, making the next iteration more accurate.
3.  Code Refactoring and Documentation: The ML and AI technical and services landscape has evolved considerably since the original build, with new approaches now available that didn't exist at the time. SPR refactored the codebase — removing unnecessary components, simplifying logic, and improving readability — so the client’s team is better positioned to maintain, extend, and eventually migrate the solution to their new platform. This reflects SPR’s Empower philosophy embedded within Evolve: not just improving the technology, but leaving the team more capable of running it independently.

This scope was achievable in four weeks in part because SPR had documented the original engagement thoroughly. Notes on Azure resource names, configuration settings, and architectural decisions remained accurate and usable years later, enabling a two-person team to move quickly without ramp-up overhead.

SPR’s approach centers on four connected phases: Explore, Empower, Engineer, and Evolve. In the original engagement, SPR engineered a durable solution with the structure and documentation to last. When the client returned, SPR helped empower the team with a cleaner, more maintainable codebase. Throughout the engagement, the focus remained on evolving the system, not just restoring it to baseline, but improving it.

The client received:

• A retrained model that outperformed its predecessor, validated against historical benchmarks
• Simplified Azure architecture that reduces ongoing maintenance complexity and improves service reliability
• A cleaner, refactored codebase ready for migration to the client's future internal ML platform
• Durable original documentation that accelerated the improvement efforts years after the initial build

SPR retrained the model on better data, simplified the architecture, and modernized the codebase. The client ended up with a stronger system than before, better prepared for what comes next.