ML Research Engineer
An ML Research Engineer builds the systems and experiments that help a research team test, compare and apply ML methods.
Written by Romain PichouPublished on
ML Research Engineer: hiring for this role?
First candidates presented within three weeks.
Definition and scope
An ML Research Engineer, also searched for as “research engineer machine learning”, combines software engineering with an understanding of models to make reliable research experiments possible. They design the systems used to run training, help test hypotheses and analyse results with researchers. Depending on the team, their work can span model code, compute infrastructure, data and evaluation. The role helps turn a testable idea into a runnable experiment and, if validated, a reusable component.
In a team training large models, they may build tools that submit jobs to a GPU cluster, track their progress and allow runs to resume after an incident. They measure compute, memory, communication or data loading constraints before changing the system. In research closer to a product, they may also compare model variants and prepare to transfer a prototype into operational use. Not every role combines all these activities: the scientific problem, scale of the experiments and team structure determine the emphasis.
The role works closely with researchers and the engineers who operate models or platforms. Before hiring, define who formulates hypotheses, who decides the evaluation criteria and who is responsible for the reliability of runs. Google DeepMind distinguishes research and engineering roles, while Anthropic describes substantial overlap between them. A job title alone does not settle this division of responsibilities.
ML Research Engineer, AI Researcher and Machine Learning Engineer: who does what?
- ML Research Engineer: makes experiments technically possible, helps conduct them and makes the systems running them reliable; they may contribute to methods depending on the team.
- AI Researcher: may take greater responsibility for formulating new hypotheses or methods, although some teams share this work with Research Engineers.
- Machine Learning Engineer: may focus more on integrating models into software or services; the boundary depends on the role and the expected handover.
Why this hire matters
The first question is what is actually slowing the work down. If researchers spend time restarting training runs, finding the parameters of an experiment or explaining a result they cannot reproduce, expertise in systems and experimentation may unblock the research. If the difficulty lies instead in choosing scientific hypotheses, the hire needs a stronger focus on developing methods. The same title can cover these needs in different proportions.
The second question concerns the quality of decisions. An isolated performance measure does not show whether a method is worth pursuing. The person must connect the hypothesis, data, protocol and results, then explain the limits of the comparison. In distributed environments, they must also distinguish a model result from a compute or data incident. A poorly defined role could lead to hiring someone strong in technical optimisation when the team expects scientific autonomy, or vice versa.
Illustrative case: a team compares two training variants and sees an apparent improvement. Before drawing a conclusion, it needs to identify the code and data versions, parameters and run conditions. If the experiments are not comparable, better experiment tracking may be a more useful next investment than another model optimisation. This example illustrates a possible decision, not an observed company outcome.
The third question is scale. When experiments use several GPUs or machines, bottlenecks may arise from compute, memory, communication between machines or data. Monitoring, recovery and checkpointing become useful depending on the duration and size of the runs. When hiring, specify the reliability expected, which problems the person can solve independently and which platform specialists will support them. Finally, moving a validated prototype into a usable component requires a decision about who will maintain it. Without that decision, a research demonstration may remain difficult to reuse.
Salaries 2026
| Level and experience | Annual gross base |
|---|---|
| Junior0 to 2 years of relevant professional experience | 48–60 k€ |
| Experienced3 to 5 years of relevant professional experience | 60–75 k€ |
| Senior6 years or more of relevant professional experience | 75–95 k€ |
Paris market ranges, 2026.
Indicative estimate for Paris and Île-de-France, in gross annual base salary, based on remuneration for neighbouring roles rather than statistics specific to this job title. The years of experience in the grid refer to relevant professional work; other regions, bonuses and equity are not quantified.
Key missions
- Design ML experiments based on a hypothesis, then compare the results to guide the work.
- Implement or adapt architectures, training methods and data processing to suit the project.
- Build distributed training components to run reliable experiments across multiple accelerators.
- Create tools to launch, track and analyse researchers’ experiments.
- Profile training runs to locate compute, memory, communication or data bottlenecks.
- Instrument runs and provide checkpointing and recovery when their scale warrants it.
- Build reliable data pipelines for training and evaluation.
- Validate training quality and document the conditions needed to reproduce its results.
- Adapt a validated prototype into a usable component when the team assigns this handover to the role.
Skills
Technical skills
- ML programming: write maintainable Python code to train and test models.
- Deep learning: understand architectures, training methods and the limits of results.
- Experimentation: connect a hypothesis, protocol, measurement and conclusion when comparing runs.
- Distributed training: choose and diagnose how work is divided across GPUs or machines.
- Profiling: measure before optimising and distinguish compute, memory, communication and data loading constraints.
- Run reliability: instrument experiments and arrange checkpointing and recovery when needed.
- Data pipelines: ensure consistent data supply and evaluation for training.
Expected qualities
- Rigour: explain what a result supports and what remains uncertain.
- Clarity: make a technical hypothesis, its measurements and its limits understandable.
- Collaboration: coordinate their work with researchers and platform or product engineers.
- Autonomy: break down an ambiguous training problem and justify the order of investigation.
- Judgement: connect research goals to compute and reliability constraints.
Common stack
Background and training
No single educational route defines this profession. In the Anthropic pre-training vacancy reviewed, an advanced degree (master’s or doctorate) is listed among the desired qualifications. The vacancy nevertheless states a bachelor’s-level degree or an equivalent combination of education, training and experience as its minimum level. Elsewhere, technical achievements and the ability to conduct experiments also provide relevant evidence. The hiring task is to establish what a candidate can build, measure and explain, without turning a qualification in one vacancy into a general requirement.
A software engineering background can lead to this role if the person understands models, can design experiments and works well with researchers. A more research-oriented background may also fit if the candidate writes robust code and understands the constraints of training systems. Roles focused mainly on GPU clusters require different experience from roles that alternate between model design, evaluation and transfer into a product. The job description should make this emphasis clear.
Useful evidence is concrete: a delivered system, research code, an open-source contribution or a substantial personal project. For each example, establish what the person actually did. Can they explain the initial hypothesis, the data and measures used, a failure they encountered and how they validated their fix? Distributed training work is particularly informative if the role requires diagnosing slowdowns or long-running incidents.
Expected autonomy follows from the responsibilities assigned. One role may ask the person to lead small research projects within a team; another may require them to solve an ambiguous infrastructure problem alone. The data in this profile do not establish a general number of years of experience. Describe instead the decisions the person will need to make without supervision and the expertise available around them.
Hiring this profile
When to hire
Hire this profile when progress depends on the team’s ability to run, compare and reproduce ML experiments, and that capability needs to be built within the team. The need may arise as experiments multiply: researchers then need tools to track parameters, code versions, metrics and results. Define which experiments the engineer must make possible and what decisions the team needs to make from their results.
If training runs span several GPUs or machines, specify the problems to solve. Does the role involve launching jobs, diagnosing slowdowns, monitoring runs, recovering from incidents or making data more reliable? These responsibilities call for distributed systems experience and close collaboration with researchers. They do not imply that every team has the same compute scale or tools.
If a validated method must then be used in software or a service, clearly assign responsibility for the handover and maintenance. The Research Engineer may prepare the prototype for this use, but some organisations assign part of the work to a product or platform team. State where their work ends and where another team takes over.
The deciding factor is the nature of the lasting bottleneck. If the team already has reliable experimental infrastructure and chiefly needs new hypotheses, a more scientifically focused research profile may better meet the need. If the main difficulty is integrating and operating a model already selected, consider an ML engineering or platform role instead. For a defined, limited problem, temporary technical expertise may be enough. If you need to search for a specialist in a narrow market, talent search for a targeted hire may be considered once the role has been defined.
Career path
A Research Engineer may deepen their expertise in distributed training, experiment design or ML systems performance. They may also take responsibility for components used by several research projects, or coordinate technical choices linking researchers and engineers. These are possible responsibilities, not an automatic progression.
Depending on their work and the organisation, a move into an AI Researcher role may give them greater responsibility for formulating hypotheses. A Machine Learning Engineer role may focus more on integrating models into a service. MLOps Engineer or Platform Engineer responsibilities may suit someone who wants to focus on the reliability and operation of infrastructure. The direction of any move depends on the work actually done in each team. Another route is to remain an expert technical contributor and tackle more difficult research or systems problems without taking on management responsibilities.
How to assess this profile
To assess this profile, start with work the candidate has actually done, then offer an exercise close to the problems of the role. Adapt your criteria to the research and infrastructure responsibilities you will assign.
1. Set criteria before interviews
Describe a problem to solve and the decisions the person will need to make. If the need centres on ML experiments, assess their ability to formulate a hypothesis, choose measures and interpret results. If distributed training dominates, add performance diagnosis, run reliability and code quality. Specify who validates scientific choices and who operates the infrastructure. Separate these responsibilities in your assessment criteria and prioritise those the role will actually carry. An expectation of universal expertise in models, data and platforms without clear priorities signals a poorly defined role.
2. Examine past work
Ask the candidate to present an experiment from hypothesis to conclusion. Clarify their contribution to the code, data, measurements and analysis. Ask what changed after an unexpected result and how someone else could reproduce the experiment. For a systems project, examine the initial problem, measurements before and after the fix, and the solution’s limits. A strong answer distinguishes what the candidate observed from what they assume. Be wary of a claimed improvement without a comparison protocol, or a team project where their contribution remains unclear.
3. Offer a practical case close to the role
Present a run that slows down or fails across several GPUs. Ask what information the candidate would gather before changing the code. They might examine metrics, data loading, memory, communication between machines and error traces. Ask them to explain how they would isolate a cause, test a fix and check that training quality has not changed. If the role focuses more on models, use a comparison of two variants under explicit test conditions instead. Reasoning and validation matter more than the name of any particular tool. An immediate optimisation without an initial measurement is a warning sign.
4. Assess collaboration and autonomy
Ask how the candidate would discuss a desired experiment with a researcher when the system imposes a constraint. Look for an explanation of the options, compute cost and consequences for the hypothesis being tested. Also establish what they would do alone when faced with an ambiguous incident and when they would consult a platform specialist. A good discussion makes the trade-offs understandable without hiding uncertainty. A warning sign is a promise of a technical solution without checking whether the measurement answers the research question.
5. Corroborate with references
With the candidate’s consent, ask people who have worked with them what they actually built and how they handled a failed experiment or training run. Compare the responsibilities described in the interview with deliverables and ways of working with others. If your company lacks the expertise to judge the technical case, involve an experienced researcher or engineer in choosing the exercise and assessing the answers. Apply the same criteria to the candidates you compare.
Frequently asked questions
How does an ML Research Engineer differ from a Machine Learning Engineer?
A Research Engineer may design the experiments and systems used to test research ideas; a Machine Learning Engineer may focus more on integrating models into software or services. The boundary varies by organisation. Compare the expected deliverables: an experimental protocol and infrastructure, or the operation of a model already selected. If the role includes both, specify who decides on methods and who maintains the component after handover.
Is a doctorate required for this role?
No, a doctorate is not an established general requirement for this profession. Some vacancies seek advanced education, while other hiring decisions value delivered systems, research code or open-source contributions. Define the capabilities the role needs: understanding models, conducting experiments and building the required system. A particular degree may still be justified by a specific project if you can explain which capability it should demonstrate.
How should we define the autonomy expected at the start of a research project?
List the decisions the person will make independently from the first experiments: choosing measures, changing training code, diagnosing incidents or organising compute resources. State which decisions remain shared with researchers and the platform team. This distinction helps separate a role in which the engineer leads small pieces of work within a team from one in which they must solve ambiguous infrastructure problems alone. It also provides a more concrete criterion than a seniority label.
How should we interpret remuneration for an ML Research Engineer?
According to this salary grid, estimated gross annual base salary ranges from €48k to €95k in Paris and Île-de-France in 2026. This estimate is based on remuneration for neighbouring roles, not direct statistics for this position. To compare an offer with the grid, consider responsibilities in research, training infrastructure and transfer into a product. Any overall package, bonuses, equity and other regions are not quantified.
What work sample can show whether a candidate makes experiments reproducible?
Ask the candidate for a record of an experiment they can present. It should connect the hypothesis to the parameters, code version, data, metrics and run artefacts. Ask them to explain how they retrieve that experiment and check whether a change produces a comparable effect. The aim is not to require a particular tracking tool: you want to understand what information their team can reuse to analyse and resume work.
Sources and method
- Google DeepMind : Careers at Google DeepMind
- Anthropic : Research Engineer/Research Scientist, Pre-training
- Together AI : Research Engineer, Large-Scale Training
- Runway : Member of Technical Staff, Research Engineer
- Pluralis Research : Research Engineer, Pre-training
- MLflow : ML Experiment Tracking
- Cognition : Research Engineer, ML Infrastructure
Related job profiles
- Artificial intelligence researcher (AI Researcher)An artificial intelligence researcher designs and evaluates new methods to address an organisation’s scientific questions.
- ML Engineer (Machine Learning Engineer)An ML Engineer designs, trains and integrates machine learning models into software used by customers or teams within the company.
- MLOps EngineerThe MLOps Engineer automates the deployment of machine learning models to production and organises monitoring for the teams that operate them.
- Platform EngineerBuilds the internal platform (tooling, environments, golden paths) that lets product teams ship fast and well, autonomously.
- Data EngineerBuilds and operates the pipelines that collect, move and expose data at scale: ingestion, warehouse, orchestration.
About the author

Co-CEO
Romain Pichou a cofondé GetPro en 2015 avec Émile Pennes. Diplômé de l'ESCP Business School, il a débuté sa carrière dans des entreprises technologiques en forte croissance (Winamax, Betclic, Lucca où il dirigeait les ventes de la suite SaaS RH, puis ContentSquare).
Chez GetPro, il est l'associé référent des recrutements Tech, IA et Produit : CTO, VP Engineering, Head of Data, direction produit. Il intervient sur les mandats de direction technique, du cadrage du besoin à l'évaluation des candidats.