Банк вопросов из реальных собеседований

How to store a production plan extracted from documents so that new reports correctly update forecast features?

Why might a model based on historical production be dramatically wrong if a company invests in a new mining method or mine expansion?

What are the differences between the W8A8 and W4A16 diagrams for the LLM application, and which indicators should be compared before choosing?

How to add text and visual features to a recommendation system without breaking the serving pipeline?

How to select and adapt a text encoder for queries like “walker crosses the road at night”?

A speech-AI pipeline needs fast analytical queries over training-data processing events. What requirements would you give DevOps before asking for ClickHouse?

How can you increase LLM serving throughput or batch size on the same GPU without buying a larger GPU?

What techniques help to obtain a stable, verifiable and safe result in the work system?

You trained and evaluated an ML model. What model files do you version, how do you package the service and how do you roll it out safely?

Which products must not be recommended in the cart, and where should the constraints be applied: candidate generation, filtering, or reranking?

The model can evaluate the price, discount, utility of the carousel or promo tags. How exactly to determine the output of the model and the subsequent product action?

A bank wants to launch a lifestyle social feed without ads. What goals, metrics and guardrails would you define before designing the recommender?

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?

Explain how LLM tool/function calling works end to end: tool schema in the prompt, model output, real tool execution and final user response.

What components transform a language model into an agent and for what tasks is the cycle of external actions justified?

What do you do when an Airflow DAG brakes, freezes, or fails to fit into a scheduled window?

What approaches are there for training a large neural network on several GPU and how do they differ?

What is a reliable pipeline for you and how to verify that it is reliable?

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?

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

The LLM-agent product already has an offline benchmark: for each change you can see whether the metric has become better or worse. How to turn the evaluation results into a cycle of improving the system, without slipping into blind automatic optimization for a noisy benchmark?

Design a system that, based on the photo and metadata of the ad, determines that different cars have appeared in the card or car history.

How do FSDP, tensor parallelism and pipeline parallelism differ when training large models?

The command changes prompts/models/rules for invoice parsing. How not to break quality with every change?