Home / Machine Learning Engineer
Machine Learning Engineer Interview Questions (2026)
ML engineer interviews still test classical fundamentals — the difference from AI engineer loops is depth on training, data, and deployment of models you own. Expect coding, ML theory, and system design rounds. These questions recur across FAANG and AI-startup loops.
1.Explain the bias-variance tradeoff and how you diagnose which one is hurting your model.
What a strong answer covers: High bias: poor performance on training data itself — the model is too simple. High variance: large gap between training and validation performance — the model memorizes. Diagnose with learning curves. Fixes differ: bias wants more capacity or features; variance wants regularization, more data, or simpler models. Tie it to a real project where you diagnosed one — the concept is table stakes, the application is the differentiator.
2.Your model performs well offline but poorly in production. What are the possible causes?
What a strong answer covers: The most important MLE question. Cover: training-serving skew (features computed differently in the two pipelines — the top real-world cause), data drift, label leakage that inflated offline metrics, feedback loops, stale features at serving time, and offline metrics that don't match the business metric. Describe how you'd instrument to distinguish them: log serving features and score them offline.
3.How do you handle severe class imbalance, like fraud detection with 0.1% positives?
What a strong answer covers: Start with the metric: accuracy is meaningless; use precision-recall AUC or cost-weighted metrics tied to business impact. Then techniques: class weights, resampling (and the risks of naive oversampling), threshold tuning on the operating point the business needs, and anomaly-detection framing when labels are scarce. Mention that getting more labeled positives often beats any algorithmic trick.
4.Design a feature store. Why do companies build them?
What a strong answer covers: The problem: features computed inconsistently between training and serving causes skew, and teams rebuild the same features. Cover offline store (training, backfills) vs online store (low-latency serving), point-in-time correctness to prevent leakage, and feature sharing/discovery. Naming the consistency problem as the core motivation shows you've felt the pain, not just read about Feast.
5.Walk me through how gradient descent works, and why Adam is usually the default optimizer.
What a strong answer covers: Cover the basics crisply: follow the negative gradient with a learning rate, minibatches for tractability. Adam adds per-parameter adaptive learning rates from first/second moment estimates, making it robust to learning-rate choice. Mention learning-rate schedules and warmup for transformer training. Depth check: they may ask why you'd ever prefer SGD (better generalization in some vision tasks, simpler behavior).
6.How would you build a recommendation system for a marketplace from zero data?
What a strong answer covers: Cold-start system design. Phases: heuristics and popularity first (with position-bias awareness), content-based similarity from item metadata and embeddings, then collaborative filtering once interaction data accrues, then a learned ranker with business objectives blended in. Cover the feedback loop trap: the system only learns about items it shows. Logging design from day one is the mark of experience.
7.Explain how transformers work well enough that I believe you could modify one.
What a strong answer covers: Beyond the diagram recitation: tokens to embeddings plus positional information, self-attention as learned weighted mixing of token representations (queries/keys/values), multi-head for different relation types, MLP blocks, residual connections and layer norm for trainable depth. Be ready for follow-ups: why attention scales quadratically, what KV-cache does at inference, what changes in a decoder-only model.
8.How do you decide when a model needs retraining, and how would you automate it?
What a strong answer covers: Triggers: performance decay on fresh labeled data (best but often delayed), input distribution drift (PSI/KL monitoring as a proxy), and scheduled retrains when labels lag. Automation: pipeline that retrains, evaluates against the incumbent on holdout and recent data, and requires a gate (automated or human) before promotion. Mention shadow deployment for validating the challenger safely.
9.You have 10M unlabeled examples and budget to label 10k. How do you spend it?
What a strong answer covers: Active learning framing: label a random seed set for an initial model and unbiased eval, then select by uncertainty and diversity (cluster embeddings, sample across clusters) rather than pure uncertainty which picks redundant edge cases. Hold out part of the budget for a clean random test set — skipping that makes your metrics unbelievable. Mention weak supervision and LLM-assisted labeling as 2026-relevant multipliers.
10.What metrics would you use for a ranking problem, and why is NDCG preferred over precision@k?
What a strong answer covers: Cover precision/recall@k, MRR for single-relevant-item tasks, and NDCG which handles graded relevance and rewards putting the best items highest — position matters, which precision@k ignores. Then the production reality: offline ranking metrics often diverge from online engagement; describe interleaving or A/B tests as the ultimate arbiter.
11.Tell me about a model you shipped end to end. What broke after launch?
What a strong answer covers: The behavioral anchor of MLE loops. Structure: problem and business metric, data work (usually the bulk), modeling choices and what you tried that failed, deployment architecture, and monitoring. The 'what broke' part is where credibility lives — real answers involve pipeline failures, drift, or an offline-online metric gap, plus what you changed. A story with zero failures reads as a story with zero production time.
12.How is your role changing now that foundation models replace many bespoke models?
What a strong answer covers: Honest and current: many tasks that needed custom models (classification, extraction, NER) now start with a prompted LLM, so MLE work shifts toward data quality, evaluation, fine-tuning open models, cost/latency optimization, and the tasks LLMs still can't do (tabular prediction, ranking at scale, anything needing calibrated probabilities). Framing this as expanded leverage rather than threat shows the mindset teams want.
Prepping for a specific job?
Use the prep engine on the homepage — describe your interview and get a tailored question set instantly. Missing a role page? Request it.
Want a human in your corner? 1-on-1 interview prep — $150
A 1-hour session where I use AI to build a prep plan for your exact interview — the role, the company, the round. Strongest for technical interviews (cloud, AI/ML, DevOps) and behavioral rounds: mock questions, answer structuring, and a follow-up question bank tailored to your job description.
Book a session