07Machine Learning on Microsoft Fabric (Data Science Experiments)
Role: Pre-Sales Solution Engineer
Executive summary
Created demos and prototypes proving end-to-end PyTorch ML workflows on Microsoft Fabric, clarifying Fabric vs Databricks positioning for FSI clients evaluating Fabric as an ML platform.
- Microsoft Fabric
- Fabric Data Science
- Spark notebooks
- AutoML
- PyTorch
- MLflow
- Azure Databricks
As Fabric introduced integrated Data Science, many Financial Services clients asked how it works as an ML platform. For pre-sales enablement, I built prototypes and demos showcasing PyTorch ML workflows on Fabric versus Azure ML and Databricks.
- Create sample Data Science experiments in Fabric (e.g., training a small PyTorch model in Fabric notebooks on Spark).
- Show experiment tracking with MLflow in the Fabric UI.
- Highlight ease of use: Fabric Warehouse as source, Spark training, model saved in OneLake.
- Compare performance and features with Databricks to articulate enterprise readiness.
I built demos importing data into a Fabric Lakehouse, doing feature engineering in a Data Engineering notebook, then training PyTorch models in Fabric's Data Science experience—e.g., a credit-risk classifier built entirely in Fabric. I measured training time, consulted product contacts on platform nuances, and prepared comparisons (Fabric's Spark engine vs Databricks Photon, cost differences). For real engagements, I tailored demos to client data and use cases as quick pilots.
The demos convincingly showed Fabric handling typical ML workloads—one pilot matched the accuracy of an existing Databricks process for a small model. They also clarified Fabric vs Databricks positioning: Fabric for integrated analytics in a unified environment, Databricks for mature MLOps—supporting a “better together” story. The prototypes strengthened competitive messaging and my ability to advise clients.
As an early explorer of Fabric's ML capabilities, I learned to navigate new-platform quirks. Being tool-agnostic yet deep in both Databricks and Fabric let me propose optimal solutions. Measuring rather than assuming—training times, integration ease—was crucial for credible guidance.