A team building a support-ticket classifier needs more than a trained model file. It records the training data and code versions, checks predictions on held-out tickets, and packages the model with the same text preprocessing used during training. A release can then be reproduced and compared with its predecessor.
After deployment, the team checks incoming data and prediction outcomes. A new product name or changed support policy may make an old model less useful even when its code still runs. Monitoring provides a reason to investigate; it does not automatically establish that retraining will fix the problem.
This matters because a machine learning system includes data pipelines and serving software as well as learned parameters. LLMOps applies related practices to language-model applications, adding concerns such as prompt versions, retrieval and tool traces. Neither label replaces task-specific evaluation or clear responsibility for a release.
Sources
- Google Cloud: MLOps, continuous delivery and automation pipelines — Explains versioned development, data and model testing, deployment, monitoring and training-serving differences.
Go deeper
- Microsoft Learn: Introduction to MLOps course
Follow a learning path on experimentation, pipelines, deployment and monitoring.