Банк вопросов из реальных собеседований

A vector search returns top-k nearest items, but all results are too similar to each other. How can you keep relevance while increasing diversity?

A deployed LTV model works acceptably on average but badly for some apps. How would you find where and why it fails?

There is an idea to determine the audio event around the user: a dog barking, opening a door or broken glass. How to check product benefits before model training?

How to choose thresholds for phishing warnings and what metrics to monitor in production?

Design an ML system that personalizes the free-delivery threshold or offers paid delivery below it. How would you define the objective, decision boundary, and first version?

You need to build cart recommendations for 10 million users and 100,000 products. How would you define the objective, constraints, and first version?

You replace brute-force nearest-neighbor search with a Qdrant HNSW index in an offline recommendation pipeline. What design and operational choices matter?

How to technically build a model that determines the event by audio: the barking of a dog, the sound of a door, broken glass and similar classes?

You need to do a search/hints in the database of articles or bank answers. Why is it wise to start with BM25/TF-IDF rather than straight away with embeddings/RAG?

What determines the high share of acceptance of preliminary tokens and why the average value should be analyzed by positions and types of requests?

It is necessary to estimate or suggest the price of the advertisement for the sale of used goods. What are the signs and basic options to offer?

What feature groups should be named in RecSys ML System Design: user, item, context and engineering features?

Why can seller traits help a pricing or ranking model in a classifier and what equity and cold start risks need to be controlled?

How do you determine whether the new model will improve the product and justify the cost of switching from the current solution?

You need to predict retention, revenue or LTV over time. How to set a task, choose a horizon and granularity, build a baseline and carry out correct time validation?

After selecting price model features or recommendations, how to determine the target variable, partition, baseline, model and launch criterion?

Why should you map the full user journey and order-fulfillment process before selecting a model for delivery pricing?

How was quality assessed: how well are you able to conduct a dialogue, answer a question, or search for the necessary documents?

Item-item recommendations may have no neighbors because a product is new or rare, the user is new, or the cart is large. Which fallback strategies would you use?

With 10,000 calls per day and many quick rejections, how would you structure the pipeline and choose metrics before running an expensive extractor?

Which real-time kitchen and courier features are available to the pricing model, and how do they differ from stable user and restaurant features?

How do you move from revenue, seller success, and buyer value to offline pricing model metrics or recommendations?

What facts from company PDF reports are useful for production forecasts, and how can you distinguish them from noisy text?

One fact about the mine appears in the annual report, presentation and call transcript. How to combine these sources into one forecasting state?