Банк вопросов из реальных собеседований

The interviewer mentions DFlash as a diffusion-style speculative decoding algorithm. How would you explain the idea, candidate verification, and its trade-offs?

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

How does function calling work and how does structured output work?

How to use the hidden profile of the student and prevent the user from pulling it out via prompt injection?

Design the end-to-end pipeline for a RAG system: data preparation, vector index ingestion and serving-time retrieval.

The CTR panel has a stream of events about impressions and clicks. How to separate the roles of Kafka, S3 and ClickHouse?

Explain the difference between a Kubernetes pod, service, deployment and node.

What does the language model’s KV cache store, why does it speed up generation, and what limitations does it create when serving multiple queries?

Why can the historical backtest of a LLM feature be unfair, even if the documents are submitted with the correct dates?

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

A deployed ML service has 300 ms latency, but the product now needs 30 ms. What do you investigate and what optimizations can you try?

Design a basic multi-head attention block in PyTorch. What tensors do Q, K and V have, when do you split heads, and what common shape mistakes should you avoid?

What baseline should I launch for a new video feed if there are no clicks or purchases on it yet?

In invoice parsing, some documents are native PDF, some are scans. How do you determine which processing path to use and what errors to expect?

The moderation model works in production. What metrics should you look at offline, online and after launch in order to control the quality and load on manual checking?

How to embed a ML reranker into an existing search if candidate generation already returns itemIds?

Article searches can be evaluated offline, but it is important for the product whether it helped the user. What online signals show this?

Why do LLMs need positional embeddings, and what implementation variants are commonly used?

What are the prefill and decode stages of LLM generation, and how do they differ?

In a screening interview, explain a production vector-search or RAG project: technologies, index, traffic model, deployment and what part you owned.

After launching the CV model at real points, errors, new dishes and new shooting conditions appear. How to build support and additional training?

A video-analytics platform wants an assistant that answers support requests using documentation and prior tickets. How would you design the RAG system?

Which mechanisms can give a Transformer positional information, and how does RoPE work?

The user writes a natural query to a large corporate database. How to match entities, abbreviations, tables and columns?