Overview
A major Australian telecommunications provider needed to move beyond ad hoc notebooks and build a production-ready machine learning capability. Innablr designed and delivered a secure, scalable Data Science and MLOps platform on Databricks — enabling the client's own team to build, deploy, and monitor churn prediction models independently, with a reusable foundation for every model that follows.
Client: Major Australian Telecommunications Provider
Technologies: Databricks, MLflow, Unity Catalog, AWS, Azure DevOps CI/CD, Infrastructure as Code (IaC), GitOps
Challenges
The client's Data Science team had the capability to build models but lacked the infrastructure to take them to production. The primary objectives of the engagement were:
- Secure Data Access: Provide governed, read-only access to production data for model training without compromising security controls.
- MLOps Infrastructure: Establish automated CI/CD pipelines for machine learning to ensure consistent and reliable model deployment.
- Scalable Compute: Provision dedicated compute resources tailored specifically for data science workloads.
Solution
The Solution
The Innablr team structured the engagement across three parallel workstreams — data access, MLOps infrastructure, and data ingestion — delivering working infrastructure at every milestone rather than a single big-bang release. By building the platform before the model, DX1 ensured that every subsequent initiative would be faster, safer, and more observable than the one before it.
Key Technical Pillars:
- Secure Sandbox & Production Read-Only Access: Provisioned a dedicated Sandbox schema and established governed, read-only access to production data via Unity Catalog, with synthetic datasets to unblock the team from day one.
- MLOps Framework: Configured automated CI/CD pipelines, dedicated compute clusters, and a full model training, evaluation, and monitoring framework using MLflow.
- EDW Gap-Fill Ingestion: Identified missing variables from the Enterprise Data Warehouse and engineered ingestion pipelines into the Databricks Data Lake to maximise model accuracy and coverage.
Key Outcomes & Results
- 30 Days to Governed Environment: The data science team was unblocked within the first month, operating in a secure, two-tier environment with full Unity Catalog governance in place.
- 90 Days to Production Model: The first churn prediction model was deployed to production via GitOps and automated MLOps pipelines — a first for the organisation.
- Reusable MLOps Platform: The infrastructure built for churn is not a one-off. Every model the team builds from here follows the same governed, automated path to production.
- Team Capability Uplift: DX1 worked embedded with the client's data science team throughout — not a handover at the end. The team owns the platform.
- Self-Service Analytics: Genie natural-language querying gives Finance and Operations direct access to answers, reducing reliance on the data team and cutting time-to-insight.