Банк вопросов из реальных собеседований

How to connect business metrics of a product with offline metrics of a recommendation model?

What are the ways to build pre-tokens used in servicing LLM and how do their cost, memory and dependence on the main architecture differ?

How does vector search differ from full-text search and why is hybrid often needed?

How to compare architectural options of a recommender system and choose a sustainable option?

How to choose FAISS, HNSW-based CPU index, Redis, Qdrant or Elasticsearch to search for the nearest embedding? What parameters and metrics should I look at?

How to choose an industrial model of price or recommendations, if the more complex version shows the best offline metric?

There is a response model for several delivery conditions. How to choose the final cost or minimum order amount?

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

Where does the shopping cart composition recommendation end and the personalized referee using the user history begin?

The couriers are assigned to the point and zone, and the pricing system does not control the schedule. How should this limitation affect the design of the ML solution?

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

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

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 the recommendations on the joint occurrence of goods: long tail, new products, bias to the popular and lack of novelty?

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

Why can a similar ring next to an already added ring be worse than a supplementary product? How to distinguish between substitutes and supplements?

The customer can remove suggested suppliers and add his own. How to explain the value of a recommender system in such a product and what metrics follow from this?

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

How to explain the acceleration of FlashAttention through the memory hierarchy of GPU, without going into the details of the implementation of the cores?

After ASR, how would you choose and evaluate an LLM that extracts branch, date and time as JSON from a call transcript?

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