Банк вопросов из реальных собеседований

How to design candidate generation and ranking for selecting suppliers for a procurement application in the corporate procurement marketplace?

There is a DSSM-like model trained in click-through sessions with the original, positive and negative merchandise. How to add text embedding products from LLM?

The model should run on a speaker or camera with a CPU and limitations on latency, privacy and power consumption. How do we design this solution?

How do you apply perception, prediction and maps from an existing autonomous driving system to search for traffic scenes?

What if historically the cost of delivery has rarely changed and there is almost no variability for learning elasticity?

Historical data is available only for suppliers who have already been shown or invited. How to understand and reduce selection bias, and how to deal with losing bids?

LLM extracts features from a PDF report: for example, a future production plan. How can you check that a feature is based on a document and not on external knowledge or guesswork?

We need to add a LLM agent to the existing product pipeline. How to design agent boundaries, tools, context, checks and monitoring?

Retriever returned top-K chunks. How to choose the final context for LLM and where is the reranker needed?

There is a domain search where generic embeddings do not work well. How to retrain the encoder and check that retrieval has become better?

There is a production forecast for mines/assets. Company reports include text, growth plans, schedules and future expectations. How can I use LLM to improve the tabular model without completely replacing it?

Legal documents are difficult to cut with a fixed window. How to build chunking for legal or enterprise RAG?

There is an auto ad base and an anchor ad. You need to find archival ads where visually another car: a different color, interior, wheels, body or angle. How to set a task and train a model?

The report said: production will grow by 20% for the year, growth will begin in the second half of the year. The model needs a quarterly forecast. What should the LLM feature return?

There is an LTV/retention forecast for cohorts. For old cohorts there are 7/30/180 days of history, and for new ones there is only an install or trial. How to make a forecast today?

There is an LTV/retention forecast for cohorts. For old cohorts there are 7/30/180 days of history, and for new ones there is only an install or trial. How to make a forecast today?

In addition to answering the free-text question, you need to show related articles. How to form them: in advance or depending on the request, and where is the reranker needed?

Each customer has his own category tree: names can be normal words, internal codes, or of varying depth. How to take such categories into account when selecting suppliers?

The casino product data is stored in storage and comes through a message queue. You need to regularly update the signs and prompt users. How to design a production pipeline?

Есть реальный продуктовый use case: customer support automation. Нужно сравнить два LLM/agent variants и выбрать, какой запускать. Как спроектировать evaluation: данные, offline metrics, human/LLM judging, system metrics и online validation?

The supplier can be introduced through past applications, profile and categories. How to build a supplier view and what problems do averaging request embeddings have?

There is a transformer model of user history recommendations. How to make sure that it does not break the online service?

How to impose category and complementarity restrictions if candidates are extracted from the ANN index by embedding?

There is a sequence of representations of the road scene beyond the time window, in which several rare events can occur simultaneously. How to set a task and build a basic solution?