Hire an MLOps engineer

An MLOps engineer sets up the path each model follows from experiment to production. They put code in Git, record runs and model versions in a registry, build CI/CD with separate development, test and production environments, and set up monitoring and retraining.

On our staffing page, the MLOps engineer is listed among the data engineering roles. If you need an MLOps specialist for a particular platform, we’ll put forward someone who knows it.

A machine learning engineer takes a single model into production: from notebook to tested code, a training pipeline and an endpoint. An MLOps engineer builds the shared infrastructure used by all models: a registry, CI/CD templates, monitoring and agreements on retraining.

In small teams, one person often does both. If your main goal is to get one model live, read about hiring a machine learning engineer.

Google describes three levels. At level 0, everything is manual: the data scientist hands a trained model to engineers, new versions are rare and predictions are not monitored. At level 1, the training pipeline is automated and the model retrains on new data. At level 2, CI/CD also tests and deploys the pipeline itself.

If your team is at level 0 and runs more than one model, moving to level 1 is a job for an MLOps engineer.

They monitor inputs for data drift, predictions for prediction drift, data quality (missing values, incorrect types and out-of-range values) and, once actual outcomes are available, the model’s performance. Each signal has a threshold that triggers an alert.

Microsoft recommends choosing thresholds together with the data scientist who knows the model, to avoid too many alerts. In Azure Machine Learning, an alert can start a retraining job via Event Grid.

In most cases, Databricks recommends promoting the code: the training code moves through testing to production and trains the model there, because that code has been reviewed and tested. Promoting a trained model makes sense when training is very expensive or difficult to repeat.

In Unity Catalog, you assign an alias such as champion to the live version. Batch jobs and endpoints follow that alias, and lineage shows which tables were used to train the model.

A freelancer you contract directly must also work independently in practice. Since 1 January 2025, the Dutch Tax and Customs Administration has resumed full enforcement of this requirement. Someone who works through your backlog under your direction can quickly fall outside that definition.

Our MLOps engineers are employed by Data Science Lab and seconded to you. Ask HR, procurement or an adviser to assess how that fits your own hiring policy.

For high-risk AI systems, the AI Act requires, among other things, logging so that outcomes can be traced, human oversight and robustness. After deployment, the provider must have a monitoring system. According to the European Commission, these rules apply to most high-risk applications from 2 December 2027.

An MLOps process that records runs and versions and monitors models helps provide that evidence. A lawyer or your compliance department can determine whether a model falls under those rules. Read more about AI governance.

If you’re not running any models yet, there’s nothing to deploy or monitor. Start with data science instead. If you have one model that needs to go live, a machine learning engineer is a better fit. If the problem lies with the data itself, address the data foundation first.

If you want the entire process delivered rather than extra hands, MLOps as a project is a better fit.