ML Engineer (Machine Learning Engineer)

A Machine Learning Engineer (ML Engineer) is a specialist in software development and machine learning. They design, train and adapt models for use in products and services. Their work connects the model's behaviour with the operation of the software that uses its results.

They select the necessary data, choose an approach suited to the problem and evaluate performance. They also develop code to make processing reproducible and collaborate on integrating the model into existing systems. The role therefore combines modelling decisions with software design, testing and documentation requirements.

ML Engineer, Data Scientist and Data Engineer: who does what?

• With the Data Scientist, clarify how experimentation, model selection and the transformation of models into usable software components are shared.
• With the Data Engineer, define the boundaries of work on data flows and the preparation of training datasets.
• With the MLOps Engineer, clarify responsibilities for the platform, deployment and monitoring in production.

These responsibilities can overlap depending on the organisation. The presence of related skills in a role does not make the professions interchangeable or automatically place the entire platform under the ML Engineer's responsibility. This profile covers software engineering specialising in machine learning. It does not describe a research position or a role dedicated to operating large language models.

The challenge is to turn a trained model into a software function whose behaviour meets the company's expectations. Model quality must be assessed against the problem to be solved. Integration also requires maintainable code, tests and coordination with the people responsible for the existing system.

Before recruiting, specify the results the application must produce and the criteria for judging them. A candidate who can improve a metric must also be able to explain what that improvement contributes to the intended use. Conversely, strong software experience alone does not establish their ability to choose and evaluate a model.

Hypothetical example: a team has a scikit-learn model whose experimental results meet its objective. However, differences between the dependency versions used in training and production can cause errors or unexpected behaviour. Before integrating the model into an application, the team must therefore identify the versions used and check that it works in the intended environment. Assigning only a new round of training to the person recruited would leave this integration difficulty without anyone explicitly responsible for it.

• Translate the business need into a modelling objective and evaluation criteria with the relevant stakeholders.
• Select and prepare the data needed to train the model.
• Design, train or adapt a model according to the problem and the available data.
• Evaluate the model's results against expected performance.
• Develop reusable, tested and documented code for model-related processing.
• Collaborate on integrating the model into existing systems.
• Contribute to monitoring, retraining and improving models with the teams responsible.

French engineering degrees (diplômes d’ingénieur) and master's degrees in computer science specialising in artificial intelligence are among the routes presented by Onisep. They provide a framework for examining a candidate's knowledge and skills, without being a hiring requirement or a guarantee of autonomy on a production system.

Useful software skills include program design, structuring reusable code and testing. In mathematics, applied statistics help connect the characteristics of the data with the learning method and the intended objective. Choosing a model and interpreting metrics require knowledge distinct from proficiency with a library.

Specialist courses and practical work can complement this background. Google notably offers courses and labs to prepare for its certification. However, certification does not replace an examination of completed software work and is not enough to establish programming ability.

GetPro structures assessment around a set of prioritised criteria. This method distinguishes what can be verified in a candidate's background from what needs further exploration in an interview, with an assessment method for each criterion. Open questions call for concrete examples. Reference checks help cross-check the responsibilities held and the skills examined.

The applications to the ML Engineer role suggested below are recommendations to adapt to the position and the expertise available in your team. They do not describe a specific protocol used by GetPro.

1. Define the criteria and assessors

Build an assessment framework around data selection, modelling, software quality and integration. For each criterion, specify which decisions the person is expected to make independently and which will receive support.

Involve a business stakeholder in examining the need and an ML specialist in the technical analysis. If your company lacks this expertise, bring in an expert capable of examining the code and model choices.

Probe answers that mention tools without explaining their usefulness in addressing the role's constraints.

2. Examine a past project

Choose a project with the candidate for which they can describe their contribution precisely. Ask them to present the objective, the data used, the experiments conducted and the results evaluated.

Examine the transition from model to software: reusable code, tests, documentation and collaboration with those responsible for the system. Distinguish their contribution from what was already available. Invite them to explain the limitations of the result and the difficulties encountered.

3. Set a task close to the role

Prepare an exercise limited to the decisions actually expected in the role. Provide the necessary information and specify which points require an explanation rather than a complete implementation.

Hypothetical example: you present an existing model and an application into which it must be integrated. Ask the candidate to specify the data to examine, the evaluations needed and the software elements to test.

Invite them to justify the choice to retain, adapt or retrain the model. Assess the coherence of their reasoning and their ability to identify missing information. A definitive conclusion reached without examining the available data or expected performance warrants further exploration.

4. Observe communication and coordination

Ask the candidate to explain a technical choice to the business stakeholder. Observe whether they make its consequences and the limitations of the evaluation understandable.

Explore how work is shared with other specialists, particularly when responsibility for integration or retraining is shared.

For a lead role, examine a situation involving technical coordination or support for other engineers. If the position includes line management, assess that responsibility separately.

5. Cross-check responsibilities held

With the candidate's consent, ask someone they previously worked with to clarify their contribution, the decisions they made and the collaboration observed. Base the discussion on the responsibilities of the prospective role and ask for concrete examples.

Compare this information with the project account and observations from the exercise. Examine any discrepancies before reaching a conclusion. Distinguish a lack of exposure to a situation from difficulty explaining it.

Bring the observations together by criterion, specifying established points, uncertainties and the support needed.