How to Hire ML Researchers: Signals, Tests and Scope

Most organisations that believe they need an ML researcher need an ML engineer. Getting that distinction right before hiring prevents an expensive mismatch that typically resolves within a year by the person leaving. This guide covers it, drawing on FISTA Solutions' AI enablement work.

Do you need a researcher or an engineer?

Situation	Role needed
Known methods apply, need production quality	Engineer
Need evaluation and measurement built	Engineer
Existing approaches genuinely fail	Researcher
Novel domain with no established methods	Researcher
Adapting published work to your data	Either, leaning engineer

Research is warranted when existing methods genuinely do not solve the problem, which is rarer than assumed. The honest test is whether you have tried the standard approaches properly.

What should you test in an interview?

Evaluation design. Ask how they would know whether an approach worked, and what would convince them it did not.

Strong researchers design experiments that can fail, and they name the failure condition before running anything. Weaker ones design experiments that confirm, which produces results that do not survive contact with production data.

Do publications matter?

They demonstrate output and peer scrutiny rather than fit for your problem. A strong record in an unrelated area predicts general capability, not applicability.

Applied problems frequently reward different instincts: working with messy data, accepting a good-enough result, and stopping. Ask about a project they stopped.

What does research need to succeed in a company?

A question, a budget, and a stopping rule.

Open-ended research without a decision attached produces interesting results nobody acts on, and the function loses support at the first budget review. Tie each effort to a decision it would change.

How do you evaluate practical instincts?

Ask what they would do with limited data, which is the normal condition outside a lab. Answers involving transfer from related tasks, careful evaluation set construction, and honest uncertainty are practical; answers assuming more data can be collected are not.

What about reproducibility?

Ask how they track experiments. Research that cannot be reproduced six months later, including by its author, is not an asset.

Experiment tracking, versioned data, and recorded configurations are the practical answers.

How does the role work with engineering?

Closely, or the work does not ship. Research handed over as a notebook and a claim rarely survives productionisation.

Ask how their work reached production and what changed in translation. Researchers who have been through that are considerably more useful than those who have not.

What about honest negative results?

Ask for one. A researcher who has concluded that an approach does not work and said so has the intellectual honesty the role requires.

Organisations that punish negative results get optimistic reports and bad decisions.

How has the field changed the role?

Substantially. Much applied work that previously required training models now involves evaluating, adapting, and constraining large pre-trained systems, which is closer to engineering and evaluation design than to model development.

Ask what they think the research questions actually are now. The answer shows whether they have adjusted. See what is continuous evaluation.

Contract, staff augmentation, or permanent hire?

Augmentation suits a defined research question with a decision attached. Permanent hiring suits organisations where novel problems arrive continuously, which is a small set.

What are the common hiring mistakes?

Hiring research where engineering was needed. Screening on publication venue. Giving research no decision to inform. And measuring on papers or experiments rather than decisions changed.

How do you onboard them well?

Give them the problem, the data as it actually is, and the decision the work should inform. Researchers given a vague mandate produce vague results.

What does good look like after 90 days?

A clearly framed question, an evaluation design that could produce a negative result, a first experiment run reproducibly, and a stated view on whether the direction is worth continuing.

When do you not need this role?

When known methods would work if applied properly, which is most of the time. Hire an engineer, measure honestly, and revisit if the ceiling is genuinely methodological.

What should be measured?

Decisions changed by research output, reproducibility of experiments, and whether conclusions held when applied to production data.

What should you do first?

Write down the decision the research would inform and what result would change it. If you cannot, the need is probably engineering.

How do you attract candidates without a research brand?

Honestly, and by being specific. Researchers choose roles for interesting problems, good data, and the freedom to publish or at least to work on something worth working on. An organisation without a research reputation competes on the first two.

Ask candidates what would make the role interesting to them, and listen for whether your problem actually is. Hiring someone who wanted a different kind of work, on the assumption that compensation will compensate, produces a departure within the year and an internal narrative that research does not work here.

How FISTA Solutions helps

FISTA Solutions builds applied AI capability through AI enablement, AI agents, and staff augmentation: the research-versus-engineering question settled before staffing, evaluation designed so results can be negative, experiments tracked and reproducible, work translated into production rather than handed over as a notebook, and effort tied to decisions it would change. The record is 150+ projects for 50+ companies across 12+ countries, with 47% average efficiency gains where measured.

To scope applied AI work, message FISTA on WhatsApp, or read hire agent engineers.