Банк вопросов из реальных собеседований

In MoE LLM, where is Mixture of Experts usually located: in which layer of Transformer and why is it done?

How to build a search architecture: from primary candidates to final rankings?

Where does a cart-based item-to-item recommendation end and a personalized recommender using user history begin?

Where should we draw the line between product backend, ML service, feature store and business rules?

Couriers are assigned to a restaurant and delivery zone, and the pricing system does not control schedules. How should this constraint shape the ML design?

The moderation model requires classes and data. How to collect labels, handle imbalances, and not mix different policies into one noisy dataset?

There are historical transactions, platform logs, more than a million suppliers and about 100 customer companies. What data to use and how does scale affect the architecture?

How do you estimate the amount of data required to recognize audio events and decide when a device should enable a more expensive processing mode?

Where to get positive/negative examples for a recommendation system and what is considered ground truth?

How to choose the number of pre-proposed tokens in speculative generation and why too long a block can slow down the system?

What are the weaknesses of co-purchase recommendations for long-tail and new products, popularity bias, and novelty?

How does a bulletin board differ from an online store in terms of pricing models, rankings, and recommendations?

Why might another similar ring be a worse cart recommendation than a product from a complementary category? How would you distinguish substitutes from complements?

What line should the stream processor write to the aggregate store for the panel?

What mechanisms would you add so important ML datasets do not disappear because of human error or operational mistakes?

How to turn documents into features for forecasting: one summary, JSON-state or an event feed?

How to roll out heavy VLM into a product where there are latency and cost restrictions?

What will the runtime pipeline look like for a new procurement application and what to do with new customers, new suppliers and rare categories?

If embeddings, scores or recommendation lists are considered offline and lie in the S3/DWH, how can these results be safely transmitted backend/serving?

How does a vision encoder turn an image into visual tokens, and how are they combined with text in a VLM?

The model is already able to predict the probability of return. How to use it in the product and where to store the signs?

The team wants to add new features or models to the ranking service. How do you do that safely?

For the article you need to show short suggest questions or tips. How to get them from the text of an article without degrading the quality of the search?

How do you know if the article search or RAG system is working well? Which offline and online metrics should I use?