The word "machine learning engineer" makes most people picture someone training models, tuning architectures, chasing accuracy. So that is what teams interview for, with questions about gradient descent and the math behind a transformer. Then they hire someone who can train a model beautifully and cannot get it running reliably in production, which is where the actual value lives. Training the model is a small slice of the job. The rest is data and engineering, and the interview usually never touches it.
Key Takeaways
- Model training is roughly 10% of the job. Data and production engineering are the rest.
- Distinguish the roles you might mean: ML engineer, data scientist, and researcher differ sharply.
- Test for getting a model into production and keeping it working, which matters far more than accuracy on a notebook.
- Python fluency is assumed; the differentiator is data pipelines and deployment judgment.
Three Different Jobs, One Blurry Title
Before you interview anyone, decide which role you actually need, because "machine learning" covers people with very different skills. A research scientist pushes the state of the art and lives in papers and experiments. A data scientist analyzes data and builds models to answer business questions. A machine learning engineer takes models and makes them run reliably at scale, owning the pipelines, the serving, and the monitoring. Most startups saying "we need an ML engineer" actually need that third person, and accidentally interview for the first.
Python sits at the center of all three, having climbed to 57.9% of developers in the Stack Overflow survey largely on the strength of data and AI work (Stack Overflow 2025). So Python fluency tells you nothing distinguishing. What separates a production ML engineer is whether they can turn a promising model into something your product depends on without it silently degrading.
The 90% Nobody Interviews For
A model that scores well in a notebook is the start of the work, not the end. Getting it into production means feeding it clean data on a schedule, serving predictions fast enough to matter, watching for the input data drifting away from what the model was trained on, and retraining before quality quietly rots. These are engineering problems, and they are where ML projects actually fail. A candidate who only wants to talk about model architecture, and lights up at hyperparameters but goes quiet on deployment and monitoring, is telling you they have lived in notebooks, not in production.
| Interview usually tests | The job actually needs |
|---|---|
| Model architecture and math | Getting the model into production |
| Accuracy on a clean dataset | Handling messy, shifting real data |
| Training a model once | Monitoring and retraining over time |
| Notebook experiments | Reliable pipelines and serving |
A Concrete Screen
Skip the whiteboard derivation of backpropagation. Instead, describe a real situation: a model your team trained scored well in testing, shipped three months ago, and its predictions have quietly gotten worse. What do you do? A strong ML engineer immediately thinks about data drift, whether the input distribution changed, whether an upstream data source shifted, whether the training data has gone stale, and how they would have caught this with monitoring before a human noticed. A weaker candidate wants to retrain with a fancier model, treating a data and monitoring problem as a modeling problem. That instinct, distrust the pipeline and the data before the architecture, is the production ML skill.
The Honest Counterpoint
If you are genuinely doing research, this advice inverts. A company pushing the frontier of what models can do needs deep modeling and math talent, and a production-focused engineer who is excellent at deployment but shallow on the science will not move that needle. The same is true very early, when you are still validating whether ML solves your problem at all and a data scientist's exploratory work matters more than production reliability. Match the hire to the stage: research needs researchers, exploration needs data scientists, and shipping a dependable ML feature needs a production ML engineer. The mistake is hiring one when you needed another.
Cost and Sourcing
ML talent is among the most expensive in the market right now, and a senior ML engineer in the US commonly runs $170 an hour or more, with genuine specialists going far higher. Nearshore in Latin America, senior ML engineers land roughly $65 to $110 an hour, still a real discount at the same seniority, with the timezone overlap that matters because ML work is tightly coupled to data and product teams (LATAM rate trends). Whichever role you are filling, screen for the production reality rather than the notebook, and read our note on staff augmentation for AI and ML teams. The data-quality instincts overlap heavily with what makes a strong data engineer. See available engineers.
Frequently Asked Questions
What is the difference between an ML engineer and a data scientist?
A data scientist analyzes data and builds models to answer questions. An ML engineer makes models run reliably in production at scale, owning pipelines, serving, and monitoring. Most startups need the engineer.
What should I test when hiring an ML engineer?
Production judgment: getting a model deployed, handling messy and shifting real data, and monitoring for drift and degradation. Model-training trivia tests a small slice of the actual job.
Why is model accuracy a weak interview signal?
Because a model scoring well on a clean dataset is the easy part. ML projects fail in production on data quality, drift, and serving, which accuracy questions never surface.
How much does a senior ML engineer cost?
In the US, commonly $170 an hour or more, higher for specialists. Nearshore in Latin America, roughly $65 to $110 an hour at the same seniority.
The Bottom Line
Hiring a machine learning engineer goes wrong when you interview for the 10% of the job that is training models and ignore the 90% that is data and production. Decide first whether you need a researcher, a data scientist, or a production ML engineer, because they are different people. Then test for the unglamorous reality, deployment, data quality, and monitoring, because that is where ML actually delivers value or quietly falls apart.
Roberto Espinoza is CEO of Ruzora, which helps US startups hire pre-vetted senior LATAM engineers, with a vetted shortlist in 72 hours. See available engineers.
