Production ML Platform Support

Kept production ML scoring models and their upstream data pipelines running, including a live TensorFlow 2.13 to 2.15 upgrade with no disruption to inference.

Context

This is the data engineering side of MLOps, not model building. Models only work in production if the pipelines feeding them do too, and if someone actually owns the dependency chain underneath them.

What I owned

I maintained and monitored production ML scoring models and the data pipelines that feed their training and scoring, and managed a full dependency upgrade chain, TensorFlow 2.13 to 2.15, for a live production scoring environment.

Why the upgrade mattered

Upgrading a dependency chain under a live inference workload without disrupting it means understanding what actually changed between versions, not just bumping a version number and hoping. The scoring environment kept serving through the upgrade.

The point

Model quality gets the attention. Whether the data feeding a model is complete, on time, and correctly shaped, and whether the runtime it scores on is actually maintained, is what determines whether a model that works in a notebook keeps working in production. That’s the part I own.