Question 1
What decision would a model support in this job, and what would you refuse to automate?
Approach. Name the decision, the label you can actually collect, and the human who still owns the call.
Machine learning engineer interview questions from a job description
See the modeling, pipeline, and production questions a machine learning engineer job description tends to produce.
Matched in your browser from the requirements in the posting and Luna’s interview-question hub. These are practice prompts, not a claim that this wording will be asked.
Requirements in the posting
Machine Learning Engineer
Ship models that a product can call, not only train.
You will:
- Train and evaluate machine learning models in Python and PyTorch
- Build the data pipeline that feeds features, and say what freshness the product needs
- Watch latency when the model is on the request path, and what you show when it fails
- Decide what not to automate, and how you know an online metric disagrees with the offline score
- Explain system design choices for the model: what is fast, and what can wait
We will ask about a model that looked good and was still the wrong thing to ship.
Ship models that a product can call, not only train.
You will:
- Train and evaluate machine learning models in Python and PyTorch
- Build the data pipeline that feeds features, and say what freshness the product needs
- Watch latency when the model is on the request path, and what you show when it fails
- Decide what not to automate, and how you know an online metric disagrees with the offline score
- Explain system design choices for the model: what is fast, and what can wait
We will ask about a model that looked good and was still the wrong thing to ship.
Likely questions
Role
Machine learning
Train and evaluate machine learning models in Python and PyTorch
Question 2
How would you know a model that looks good offline is hurting users?
Approach. Say the online check and the slice — a group of users or inputs — you would watch first.
Question 3
Walk through a training or feature change that did not survive contact with production.
Approach. Keep the setup short. End with the metric that disagreed with the offline score and what you did.
System design
System design
Explain system design choices for the model: what is fast, and what can wait
Question 4
Design the system behind the main workflow in this posting. What has to be fast, and what can wait?
Approach. Users and API first. Name one bottleneck and the piece you will not build in the interview.
Question 5
How would you estimate capacity before picking a database?
Approach. Use rough numbers, label every assumption, and say which number would change the design.
Question 6
What do you shed first when traffic spikes, and which user promise do you keep?
Approach. Name the load you drop and the promise tied to the product in the posting.
System design
Data pipelines
Build the data pipeline that feeds features, and say what freshness the product needs
Question 7
Design the pipeline that feeds the decision this posting cares about. What is the grain and the freshness?
Approach. Source, grain, and what a late or partial run should do. Say who notices.
Question 8
How do you backfill without quietly rewriting history people already used?
Approach. Say what is immutable, what you version, and how a reader knows which run they are on.
Coding
Python
Train and evaluate machine learning models in Python and PyTorch
Question 9
How would you structure a Python module for the core task in this posting so someone else can change it?
Approach. Name the function boundary and the test you would write before adding libraries.
Question 10
Tell me about a time a Python script or service was correct and still too slow.
Approach. Say what you measured, the change, and the cost you left in place.
What to prepare
- Prepare one model story: the decision it informed, the metric, and where it was wrong.
- Narrate one design in sixty seconds before you draw any boxes.
- Be ready to describe one pipeline: source, grain, freshness, and what happens when it is late.
- Solve one small task in Python while saying the cost of the obvious version.
- Mark the three requirements you cannot yet support with a story.
- Say each answer out loud once, in under two minutes.
Machine learning engineer postings split into research flavor and production flavor. Training, PyTorch, and metrics lean toward “when is the offline score lying?” Pipelines, serving, and latency lean toward system design: freshness, fallback, and what you will not automate. Python is the shared screen.
If the posting names a user-facing decision, they will ask what a human still approves. If it names a pipeline, they will ask what a late run should do.
Bring one model that failed in production and one serving constraint. The sample is a production ML role, not a pure research seat.
Sample job description
The posting below is already in the tool. The questions are the ones its requirements point to. Replace it with a posting you are actually interviewing for.
Related: Nvidia role-specific questions · OpenAI interview questions · Python coding speed