Core engineering
Hire Machine Learning Engineers
Engineers who take a model from a notebook to a service that stays accurate after the data drifts.
A machine learning engineer trains, fine-tunes, evaluates and deploys models as production services. Hire one when you need reproducible pipelines, honest evaluation and monitored deployment rather than a one-off notebook result. Expect $85 to $165 per hour depending on seniority and infrastructure scope.
What this role actually does
The gap between a model that scores well in a notebook and a model that holds up in production is mostly engineering. Training data has to be versioned so results are reproducible. Features computed at training time must match those computed at serving time, or the model silently degrades. Evaluation has to reflect the decision the business actually makes rather than a convenient proxy metric.
A machine learning engineer owns that whole path. In practice, most of their time goes to data pipelines, evaluation design and deployment reliability rather than to model architecture, which is usually the least novel part of the problem.
Core skills to look for
- PyTorch, and increasingly parameter-efficient fine-tuning of foundation models
- Experiment tracking and reproducibility with MLflow or an equivalent
- Feature engineering with strict train-serve parity
- Evaluation design: metric selection, slicing, and offline-to-online correlation
- Deployment: batch, real-time and streaming inference paths
- Drift detection and retraining triggers
- Distributed training and cost-aware GPU scheduling
- Solid software engineering — this is what separates an ML engineer from a data scientist
When you need to hire this role
- A promising notebook model has stalled on the way to production
- Model quality has quietly degraded and nobody noticed until a customer did
- Offline metrics look strong but online results disappoint
- Training runs cannot be reproduced from six months ago
- You want to fine-tune a foundation model rather than prompt one
How to assess candidates for this role
Ask about a train-serve skew incident
Anyone who has shipped models has one. The specifics reveal whether they built the serving path or inherited it.
Probe evaluation design
Ask how they would choose a metric for an imbalanced, business-critical decision. Look for slicing and cost-weighted reasoning, not accuracy.
Check reproducibility habits
Ask what it takes to reproduce a result from last quarter. Strong answers involve pinned data versions, seeds and tracked configuration.
Test deployment reality
Ask how they detect drift. Vague answers about retraining monthly indicate a notebook background rather than production ownership.
Typical rate bands
| Level | Experience | Typical rate | What they own |
|---|---|---|---|
| Mid-level | 4-6 years | $85-115/hr | Owns model training and evaluation within an existing platform |
| Senior | 7-10 years | $115-145/hr | Owns the full pipeline from data to monitored deployment |
| Staff / Principal | 11+ years | $145-165/hr | Designs the ML platform, evaluation standards and cost model |
Frequently asked questions
What does a machine learning engineer do?
They own the path from data to a monitored production model: building reproducible training pipelines, designing evaluation that reflects the real decision, deploying inference services, and detecting drift. Most of the work is engineering, not model architecture.
What is the difference between a data scientist and an ML engineer?
A data scientist answers questions and builds models to explain or predict. An ML engineer makes models run reliably in production. The overlap is real, but only one of the two is usually accountable when a deployed model degrades at 3am.
Should we fine-tune a model or use prompting?
Prompting first, almost always. Fine-tuning is worth it when you have consistent task structure, enough labelled examples, and a measured ceiling on what prompting achieves. A good candidate will ask for that evidence before recommending either.
How much does an ML engineer cost?
Typical vetted contract rates are $85 to $115 per hour at mid-level, $115 to $145 at senior and $145 to $165 for staff and principal, depending on whether infrastructure and platform work is in scope.
How do we tell if a model is actually working in production?
You need monitored online metrics tied to the business decision, not just offline scores. Ask candidates how they would correlate the two; the answer separates people who have run production models from people who have only trained them.