Hire an MLOps engineer

• Builds CI/CD for models: training pipelines, tests, registries, automated deployment.
• Versions data, features and models so any prediction can be traced and reproduced.
• Serves models (batch, real-time, streaming) with sensible latency and cost.
• Monitors data drift, model performance and cost; sets up alerts and retraining triggers.
• Runs the platform: Kubernetes, feature stores, experiment tracking, GPU scheduling.
• For LLM features: model gateways, prompt and version management, evaluation sets, guardrails, tracing and cost control.

• Data scientists hand over notebooks and deployments take weeks.
• No one can say which model version made a prediction.
• Model performance degrades and you find out from customers.
• You're launching LLM features and need evaluation, monitoring and cost control.
• Cloud or GPU spend on ML is growing without anyone owning it. If you have no models yet, hire a data scientist or ML engineer first.

• Has taken models from notebook to production and kept them there.
• Treats models like software: tests, reviews, rollbacks, reproducibility.
• Knows when a simple batch job beats a real-time service.
• Can explain monitoring in business terms: what alert fires, and who acts.

• "A model's accuracy dropped over three months but no code changed. Walk me through it." Strong: data drift vs concept drift, compares feature distributions, checks upstream pipeline changes, retrain policy, alerting.
• "Design deployment for a model that needs predictions in under 100 ms." Strong: online feature store or cached features, model packaging, autoscaling, canary or shadow deployment, latency budget.
• "How do you reproduce a prediction from six months ago?" Strong: versioned data snapshot, feature definitions, model registry entry, environment.
• "How would you evaluate an LLM feature before and after launch?" Strong: labelled evaluation set, regression tests on prompt or model changes, tracing, human review, cost per request.
• "Kubeflow, MLflow, SageMaker, Vertex AI or Databricks: how do you choose?" Strong: team skills and existing cloud first; avoids building a platform bigger than the team.

• Only tool names, no production incident stories.
• No answer to "how do you know the model is still right?"
• Builds a full platform for one model.

MLflow · Kubeflow · SageMaker · Vertex AI · Azure ML · Databricks · Kubernetes · Docker · Terraform · Airflow / Dagster · Feast or a managed feature store · Prometheus / Grafana · Evidently or similar for drift · vLLM / model gateways · LLM tracing and evaluation tools.

State the models you run today and how they deploy now, the 90-day outcome (for example "automated deployment and drift monitoring for our three production models"), the stack, on-call expectations, engagement type, overlap hours and pay range. Use the data engineer job description template as a base and swap the stack and outcomes.

Option 1: post a role. Post your role free: permanent, contract or fractional; remote or on-site; the country and your overlap hours. See applicants ranked by fit. Vetted profiles show a vetting report from Carbon's real-world data experience test (scenario-based, not a coding test), a working-style profile and a short video intro. Interview with a guide tailored to the role and the candidate, then agree the rate, draft the contract and sign on the platform. Option 2: test someone you found. Already have a candidate? Test a candidate you found on Carbon before you make an offer. ID checks are available on Team and Scale.