Question 1

Walk me through the most complex model you have shipped to production.

Accepted Answer

Two minutes. The user-facing decision, the modeling choice, the deployment shape, the metric that moved.

Question 2

Walk me through gradient descent and explain why we use mini-batches in practice.

Accepted Answer

Bar for any DL role. Cover memory, noise as regularization, and parallelism on GPUs.

Question 3

Explain the difference between L1 and L2 regularization with one concrete example.

Accepted Answer

Sparsity vs smoothness. Tie L1 to feature selection and L2 to coefficient stability.

Question 4

How do you detect data drift in production?

Accepted Answer

Tests practical MLOps. Cover PSI, KS test, prediction-distribution monitoring, and the action threshold.

Question 5

You ship a fraud model. The recall on the new fraud pattern is 12%. What do you do?

Accepted Answer

Practical question. Cover error analysis, label quality, feature gaps, retraining cadence, and a non-ML hotfix in parallel.

Question 6

Implement softmax in PyTorch by hand.

Accepted Answer

Tests numerical-stability awareness. Subtract the max before exp, or you fail with overflow.

Question 7

Write a function that computes per-class precision and recall from y_true and y_pred without using sklearn.

Accepted Answer

Bar for any ML role. The math should be muscle memory.

Question 8

Design a recommendation system for a marketplace with 5M items and 50ms p99 serving budget.

Accepted Answer

Cover candidate generation, ranking, embedding store, cold start, and the offline-online evaluation pipeline.

Question 9

Tell me about a model that performed great offline and badly in production.

Accepted Answer

Specific story. Distribution shift, feedback loops, label leakage are all valid root causes.

Question 10

How do you decide whether a problem needs ML at all?

Accepted Answer

Tests judgment. Strong answers cite when a heuristic, a SQL rule, or a config switch is enough.

ML Engineer Interview Prep — modeling + production systems — for engineers in Tashkent and Central Asia

Core skills tested

Salary ranges in Tashkent (2026)

What you will actually be asked

Walk me through the most complex model you have shipped to production.

Walk me through gradient descent and explain why we use mini-batches in practice.

Explain the difference between L1 and L2 regularization with one concrete example.

How do you detect data drift in production?

You ship a fraud model. The recall on the new fraud pattern is 12%. What do you do?

Implement softmax in PyTorch by hand.

Write a function that computes per-class precision and recall from y_true and y_pred without using sklearn.

Design a recommendation system for a marketplace with 5M items and 50ms p99 serving budget.

Tell me about a model that performed great offline and badly in production.

How do you decide whether a problem needs ML at all?

Practice these questions out loud — for free

Core skills tested

Salary ranges in Tashkent (2026)

What you will actually be asked

Walk me through the most complex model you have shipped to production.

Walk me through gradient descent and explain why we use mini-batches in practice.

Explain the difference between L1 and L2 regularization with one concrete example.

How do you detect data drift in production?

You ship a fraud model. The recall on the new fraud pattern is 12%. What do you do?

Implement softmax in PyTorch by hand.

Write a function that computes per-class precision and recall from y_true and y_pred without using sklearn.

Design a recommendation system for a marketplace with 5M items and 50ms p99 serving budget.

Tell me about a model that performed great offline and badly in production.

How do you decide whether a problem needs ML at all?

Practice these questions out loud — for free

Related reading