ML Engineer technical round: models, code and production

The ML Engineer role sits between data science and software engineering, and the round reflects that. AI / Machine Learning Jobs pitched as ML engineer rather than research scientist almost always weigh this engineering side heavily.

• ML fundamentals: bias-variance, regularisation, evaluation metrics, and overfitting diagnosis.
• Coding: data-structure and algorithm problems, plus writing clean, testable Python.
• Model debugging: why a model underperforms and how you would fix it.
• Production concerns: serving, latency, data pipelines and monitoring.

The signal they want is an engineer who understands models deeply enough to deploy and maintain them, not just train them in a notebook.

Practise these recurring shapes across both ML and engineering.

'Your model does well in training but poorly in production. Why?'

They probe overfitting, data drift, train-serve skew and leakage.

'Implement this transformation efficiently.'

A coding task that tests clean, correct Python and complexity awareness.

'How would you reduce inference latency without a big accuracy loss?'

This tests practical trade-offs like quantisation, batching or a lighter model.

Question: 'A model performs worse in production than in offline evaluation. How do you investigate?'

Strong: 'I would first check for train-serve skew, whether features are computed identically in both places, then look for data drift between the training set and live traffic, and for label leakage that inflated offline metrics. I would compare the input distributions and validate the serving pipeline before touching the model itself. Often the model is fine and the pipeline is the problem.' It is systematic and production-aware.

Weak: 'I would retrain the model with more data.' That skips diagnosis entirely and may not address the real cause, which is frequently a pipeline or skew issue, not the model. It shows a notebook mindset rather than an engineering one.

Balance ML reasoning with genuine software engineering practice.

• Keep coding sharp: data structures, algorithms and clean, tested Python.
• Be able to diagnose overfitting, drift, skew and leakage out loud.
• Understand serving, latency trade-offs and monitoring, not just training.

Because interviewers push on both correctness and trade-offs, rehearse explaining your reasoning aloud. A free AI voice mock interview on InterviewPrep builds an ML engineering mock from your CV and a real job description, then scores your reasoning delivery, pace and filler words, so you explain model and production decisions clearly under questioning.