Банк вопросов из реальных собеседований

What does the Central Limit Theorem say and why is it important in statistics and A/B testing?

How are the central limit theorem, A/B test design, and MDE related?

How to explain the p-value without calling it the probability of the null hypothesis being true?

In dating or matching-product top profiles receive the lion’s share of impressions, and the rest dissolve. How to diagnose and mitigate this bias without killing engagement?

How to check that the new recommendation feed works, and how to understand how long to keep the A/B test?

How do you tell if two sets of photos show the same car or substantially different cars?

How would you build positives and negatives for training a similar-items model, and what loss would you use?

How do you design an A/B test, estimate sample size/MDE, and handle cases where treatment and control users are not independent, such as drivers and passengers in a marketplace?

In the marketplace there are free and paid ads. How to give paid ads additional exposure without compromising relevance and fairness for other sellers?

How can a function be judged based on a language model or agent if its usual accuracy does not reflect its benefits to the user and business?

During the night, you need to process about 10 thousand pages of a hundred banks on one CPU server, and the data cannot be taken out. On what occasions to spend an expensive local language model?

Speech-to-text output is lowercase words separated by spaces, without punctuation. Design a model that restores punctuation and capitalization without rewriting the text.

The system compares the characteristics of the car from the ad with photos. Fraud is rare, and erroneous blocking harms the seller. How to train the model, test it and select the automatic failure threshold?

How to calculate the share of outgoing turnover associated with suspicious counterparties from statements of different banks and check the correctness of the denominator?

A new carousel or promo tag gave an increase in metrics. How do you verify that it was caused by model relevance rather than the fact that a new interface element appeared?

Missing a rare critical event is more expensive than a false alarm. How to choose model metrics and operating thresholds?

How do you train an embedding model so that a photo search returns product-relevant results rather than just visually similar images?

The product now has many client apps. How would you decide between one global LTV model, per-app models, clustered models or app embeddings?

You train a boosting-ranker for recommendations on clicks and bundles of images. How to collect a dataset, make a validation split and not retrain on popular products and old shows?

The model learns from the purchase target and picks up cheap products with discounts. How to rank so that you earn more?

A human reviewer and an automatic checker each output a list of found errors. How do you evaluate the checker?

A retail video analytics model should flag suspicious behavior, but humans do not fully agree on what “suspicious” means. How would you define success and evaluate whether the system is doing a good job?

Как оценить качество VLM, которая генерирует описание изображения для пользователя или downstream поиска?

There may be several dishes on the plate, they mix and overlap each other. Why don't segmentation or metric learning solve the problem automatically?