MLOps and data science at nib health insurance
nib is a health and medical insurance provider in Australia with approximately 1,500 employees and over 1.6 million members. The data science practice in nib sits in different business domains, such as marketing analytics, digital sales, customer experience and operation excellence.
- Review and assess the existing ML production system for various models.
- Establish standard ML patterns and workflows with MLOps best practices to incorporate experiment tracking in model development and to enable reproducible and automated model re-training.
- Develop ML deployment templates to unify model development and development into the same workflow, and to introduce continuous integration and continuous delivery (CI/CD) into the process of putting ML models into production.
- Standardise the development workflow of data science practice with MLOps best practices and bring data scientists in different business domains into a unified platform.
- “Eliiza Machine Learning Maturity Assessment” framework was used to review and evaluate the existing data science practice in nib, and output recommendations and priorities to uplift the MLOps maturity of nib’s current data science practice
- Eliiza developed a model development template, consisting of five main stages (exploratory data analysis, feature engineering, model training, model evaluation model inference), which was embedded with:
- Data Versioning
- Data Lineage
- Experiment Tracking
- Model Evaluation
- Model Release
- Eliiza developed a model deployment template as a reusable pattern for deploying ML models into production. It includes:
- Versioning Scheme for artifact lineage
- CI / CD Pipeline for automated deployment of pipelines and model inference services
- Eliiza created both a model development workflow template and a model deployment process template to unify and accelerate data science practice in nib with best practices and consistent standards.
- The data science teams across the business in nib can have a standard workflow of development and a pathway to put models into production, where business value can be delivered.
- The development time of data science can be shortened while the deployment process of data science models into production has been accelerated.