Context
At Mercy I was the technical lead for a 10-person team of senior and lead data scientists.
The team built models in individual notebooks, without shared version control or a standard path to production.
What I built
- A documented, tested MLOps process in Azure Databricks with MLflow and PySpark, including production scripts for the feature store, inference, retraining, and validation.
- Mercy's first real-time models in Azure Databricks, integrated with the EPIC EMR.
- Four models in production through Databricks Workflows: Mercy Works on Demand, Customer Retention, Clinic Volume Forecast, and Gross Financial Forecast, with output tables in ADLS.
- NLP and entity extraction on unstructured clinical text using John Snow Labs embeddings and Azure Text Analytics for Health.
- A Power BI model validation dashboard built on Azure Synapse.
- Team standards: Azure DevOps version control with mandatory peer review, model development standards, and a support group for production code.
Results
- Delivered the organization's first MLOps pipeline and its first real-time models integrated with EPIC.
- Put four models into production through Databricks Workflows.
- Moved the team from individual notebooks to version control with mandatory peer review.