Банк вопросов из реальных собеседований

How to build a target for a product reranker if there are logs of impressions, clicks, carts and purchases?

The product manager offers to improve the price, recommendations or promotional ad block. What questions and metrics should be agreed upon before choosing a model?

You are given an uncertain research-heavy ML project that eventually must be shipped as a working model file. How do you decompose the work and communicate progress?

We calculated LTV, for example 37.37. The marketer asks how much this number can be trusted, because the purchase of traffic depends on it. How to answer?

There is only a history of user interactions with tracks. How to build the first recommendation system?

Looking beyond in-app product events, what signals should you collect from stakeholders to improve your forecast or recommendation system?

The player has just entered the casino product. What signs can be collected in the first days to distinguish a potential VIP from an ordinary player?

When would you use a pure collaborative ALS or matrix-factorization baseline for a social feed, and what are its limitations?

Where to get labels for phishing detection and how to avoid getting into the feedback loop after triggering warnings?

What implicit signals can be used instead of explicit estimates, and what biases do each of them have?

What is your view on using modern GenAI or vibe-coding tools for software and ML work, and where do they fail today?

How would you evaluate and improve a summarization service if user feedback is sparse or unavailable?

Describe the full cycle of ML tasks: from setting and data to rollout, result acceptance and monitoring.

If you train on feedback from the previous recommender, what biases can appear and how can you reduce them?

Design a similar-items recommender for 1M items when the current collaborative model fails on cold-start items and misses semantic similarity.

Cross-encoder or learning-to-rank reranker can be trained not only using manual marking. What online signals are useful for article searches?

Can the next model version train on data collected by the current delivery-pricing policy? Which biases and risks does that create?

After a new pricing policy launches, how can you decide whether its logs are safe to use for training the next version?

Модель редких событий вышла в production. Как построить feedback loop: мониторинг, data drift, retraining и регулярную разметку?