Data science for
better decisions.

I developed the spend model using feature engineering and hyperparameter tuning. SHAP explainability helped teams understand what drove the predictions and assess them in the context of offer decisions.

Compared with the previous system, overall prediction error, measured by NRMSE, fell by 10%. The model also used 77% fewer input variables, simplifying the prediction system without sacrificing performance.

Model stability extended from 30 to 360 days, a twelvefold improvement that removed the need for frequent refreshes and their associated costs. Reducing underprediction added $20M in annual value within the wider offer ecosystem.

The challenge was not simply to remove duplicate variables. Some could be reconstructed from shared base variables, but we needed evidence that the new logic preserved the meaning and behaviour of the originals.

Working across teams, I led the standardisation of variable definitions and shared sources of record. We compared reconstructed variables against their source-based originals, then checked that replacing them kept downstream model results within defined tolerances.

With a strict delivery deadline, I walked leadership through the risks of proceeding without these checks and secured additional time for validation. The broader initiative also included migrating model infrastructure to Google Cloud, with source validation and continuity checks while the models remained operational.

The workflow brings together serial validation, component returns, and work-order updates that previously required repeated manual coordination. Built with Oracle AI Agent Studio, it supports concurrent processing across serials while retaining an approval gate before operational changes proceed.

I developed a Java 21 and Micronaut service to replace legacy application logic with APIs, preserving the existing operational rules. Parameterized Oracle SQL, OAuth, and OCI Vault-backed configuration supported the underlying data access and secure integrations.

A consolidated completion email gives users a summary of the batch results. The estimated turnaround improvement compares the previous one-at-a-time process with the concurrent batch workflow; it is not a claim that each individual serial processes 90% faster.