Банк вопросов из реальных собеседований

A production service already has data, but you need to change the database schema. Describe a safe migration process.

It is necessary to forecast quarterly production by mine. What features and baseline model should be built before the LLM layer?

Explain at a high level how TensorRT or similar inference optimizers speed up neural networks, and why INT8 quantization usually needs calibration.

What is test-time scaling for an LLM, and how can you improve an inference result without changing model weights?

For international search, you can translate an existing description or generate a new one in the target language. How to compare approaches?

Why do VAD and diarization matter in an operator-client call pipeline, and how would you use them before ASR and extraction?

How would you integrate VLMs, image search, and a chat assistant into a real-estate search product so that they complement the core retrieval/ranking system instead of replacing it?

How does the model generate text token by token and what work saves the cache of keys and attention values?

A product transcribes medical consultations and uses an LLM for summarization and clinical notes, but the general model confuses medical terms. How would you improve domain understanding?

What embedding architecture did you build for RAG: a regular retrieval pipeline or something more complex?

We need to deploy a text moderation service on BERT/DistilBERT. How to design input/output, policy layer, thresholds and routing actions?

How to decompose the real-time CTR panel into event ingestion, streaming aggregation, storage and API?

You need to build a system that searches for internal documents and helps answer questions. Which Pipeline should I design?

How are the Transformer encoder and decoder, self-awareness mechanism, Q/K/V, positional coding, and how do the GPT and BERT architectures differ from each other?

What blocks does a language model agent consist of, where is its state stored, and how does it safely invoke external tools?

How does a bidirectional BERT-style encoder differ from a causal GPT-style decoder, and why are they suited to different tasks?

You need to build cart recommendations for 10 million users and 100,000 products. How would you build a simple baseline from co-purchase data?

How to safely roll out a new version of the ONNX model in production: what checks to do before release, how to enable traffic, what to monitor and how to quickly rollback?

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 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?

Given a response model for several delivery offers, how would you choose the final fee or free-delivery threshold?