Situation

Approval of dental prosthesis prescriptions at statutory health insurers required manual review by claims processors. High case volumes led to long processing times and inconsistent decisions. Predictive signals in the application data indicating clear approvability remained unexploited.

Task

Develop a scalable ML platform that predicts approval probability for dental prosthesis claims based on historical decision data, enabling rule-based straight-through processing — while ensuring compliance with data privacy regulations and the EU AI Act.

Action

- Implemented an automated, pseudonymized training data pipeline from SAP BW/4HANA to S3 (Parquet format) using dlt and SQLAlchemy.

- Built a microservice architecture on Kubernetes with a Champion/Challenger pattern for parallel model deployment and shadow testing.

- Developed a real-time inference service with a synchronous REST interface and a Kubernetes operator for automated model lifecycle management.

- Integrated an MLFlow-based model registry with versioning, AES-RSA-encrypted storage, and AI Act-compliant monitoring in PostgreSQL.

- Deployed and operated the platform via Helm Charts with an ArgoCD-based GitOps workflow.

Result

Successfully delivered the platform to multiple AOK health insurers, processing dental prosthesis approval claims in real time. The system enables automated approval of clear-cut cases (straight-through processing) and prioritizes complex cases for manual review — in full compliance with data privacy and EU AI Act requirements.