Companies Candidates

Machine Learning Engineer

They decide when a model deserves to be deployed, what limits should frame its use, and what trade-off to accept between performance, reliability, cost and real-world impact.

The role at a glance

A Machine Learning Engineer neither simply trains a model nor simply exposes a prediction behind a service. The job is to turn a statistical capability into a usable system: the data has to represent the real problem, the evaluation has to measure the decision that matters, the preparation chain has to stay consistent between training and production, and the model's behaviour has to be tracked over time.

The difference between someone executing and someone genuinely holding the role appears when the results are convincing but the system is not. Strong offline performance can come from a misleading protocol. A technically correct model can push risk onto users, reproduce past decisions, or sit unused because it was never woven into the actual work.

The job therefore runs on permanent tensions: improve a metric or preserve robustness, serve in real time or precompute, automate or keep a human check. The assessment looks at the ability to make those tensions explicit and to decide on the total cost of the system, not just the theoretical quality of the model.

Market benchmarks

On the French market, the titles Machine Learning Engineer, AI Engineer, Applied Scientist and AI engineer cover very different realities. Some roles centre on modelling, others on industrialisation, operations, or designing systems built around generative models.

Demand remains sustained, but people able to cover the whole chain are fewer than those who can train a model in a controlled environment. Genuinely senior profiles stand out less for the range of architectures they know than for the hard calls they have already had to make.

Context

In an IT services firm, the assessment focuses on getting into an inherited estate of data and models quickly, telling real guarantees from presented results, and making the limits of a contractual scope explicit. The Machine Learning Engineer sometimes has to take over undocumented datasets or models with no traceability, while preparing a handover the client can actually work with.

What separates the levels

Level 1

Junior

A Junior can prepare data, train a model within an established framework and check simple results. They apply a protocol they have been given, read the expected metrics and contribute to an existing chain. When the data is ambiguous or offline performance contradicts production, they still need someone to frame it.

Market benchmark Rarely entrusted at this level

Click to read

Level 2

Mid-level

A Mid-level engineer is autonomous on a common use case. They can compare approaches, check data quality, build a coherent evaluation, diagnose a degradation and contribute to going live. Their usual limit: they improve the model properly without always factoring in the running cost or whether users can actually act on the output.

Permanent salary · Paris region 65 – 85 k€
Freelance day rate · Paris 560 – 720 €/day

Click to read

Level 3

Senior

A Senior engineer connects the data, the model, the service and the business decision. They question whether the features are genuinely available, what each type of error costs, whether the evaluation is representative, whether training and inference agree, and how drift and feedback loops behave. They know how to restrict a use case or force an abstention.

Permanent salary · Paris region 85 – 105 k€
Freelance day rate · Paris 720 – 900 €/day

Click to read

Level 4

Expert

An Expert is not the person who knows more models or always squeezes out the best metric. It is the person who settles a structural trade-off while owning what the decision costs: giving up a performance gain to cut latency, suspending a visible but under-evaluated model, or keeping part of the process in human hands.

Permanent salary · Paris region 105 – 120 k€
Freelance day rate · Paris 900 – 1,100 €/day

Click to read

The skills we assess

05
Skill categories
assessed in real situations
Hover a category to see what it measures — and why it carries weight
Data & evaluation protocol
Highest weighting
What it measures

The quality of the training set, whether the features are genuinely available, the labelling and the protocol together determine whether the performance you observe corresponds to the problem you will meet in production. A mistake here can make everything downstream misleading.

Modelling & diagnosis
Telling drift from defect
What it measures

The Machine Learning Engineer has to pick a proportionate level of complexity, make sense of unexpected behaviour, and tell drift from a data defect, a feedback loop or a badly specified problem.

Industrialisation & operations
Not a final step
What it measures

A model has to be reproducible, versioned, consistent between training and serving, monitored and maintainable. Going live is not a final step: it is part of designing the system.

Accountability & compliance
A skill of the role
What it measures

Models can affect people, draw on data whose rights are unclear, or produce decisions that are hard to contest. This weighting means treating those subjects as skills of the role, not as a late external sign-off.

Framing & communication
Recognising when a model isn't the answer
What it measures

A good practitioner has to be able to reframe a request, explain the limits, and recognise when a rule or a change of process answers the need better than a model would.

The method in action

The method in action

Y symbol
Annotated interview excerpt

The scorecard at a glance — hover an axis

Data & artificial intelligence

Other roles in this family

Data Engineer Analytics Engineer AI Engineer MLOps / ML Platform Engineer