Банк вопросов из реальных собеседований

Why can key/value vectors from previous tokens be reused during decode, and which factors determine KV cache size?

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?

What problems arise when using long context in LLM and what approaches do they address?

How do MQA, GQA, and MLA differ from multi-head attention, and how do they reduce the KV cache during decode?

How does MoE differ from a dense model, how does the router select experts, and which systems challenges arise at inference time?

How would you build item embeddings from text, images/video and categorical/numerical attributes under real serving constraints?

The ranking model is ready. How to put it into production: offline batch or online inference?

In production RAG there is FastAPI, vector DB, ranker service, MLflow, Docker and self-hosted LLM. How to describe the request path and service areas of responsibility?

We need to build a support bot for a fintech application. What components are needed and how can you reduce the risk of answering incorrectly?

Как бы ты сделал retrieval-augmented generation для короткого factual snippet в поисковой выдаче?

How to build a RAG/search system if the case is similar in scale to a large web search?

How do you design an assistant who answers the current lesson but doesn’t disclose future material?

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?

How does speculative decoding accelerate inference, and what is distinctive about EAGLE-style methods?

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?

How to transfer Vision Transformer from individual images to a time sequence without making archive indexing too expensive?

You need to design a product: the user gives a text problem, the system makes a presentation with slides, tables and pictures. How to build a pipeline?

Нужно спроектировать AI-native продукт, который по запросу пользователя генерирует качественные motion graphics. Как выбрать между pipeline и fully agentic архитектурой, как встроить human-in-the-loop evaluation и как управлять trade-off между quality, consistency и latency?

How to select a time window and combine frame views into one or more vectors for search?

There are several GPU workers in LLM serving. Why naive round-robin routing may be ineffective, and how to design an adaptive routing layer that takes into account GPU headroom and KV-cache reuse?

We need to build a system that extracts useful fields from PDF invoices from different suppliers. Which architecture should you choose?

How to check online a new model for selecting security questions if an error could miss a fraudster or block a client?

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

After gradient boosting, which neural network architecture can be tested for sequential data and on what properties of the task depends the choice?