Банк вопросов из реальных собеседований

The dataset has missing values and noisy features. How to systematically process them before training and in production?

For realtime CTR dashboard, you need to describe Kafka/event log. What event schema is needed and by what key to partition?

A legacy ML platform has accumulated technical debt. How do you decide whether to rewrite it or keep improving it incrementally?

In a microservice system, services communicate through API and events. How to document and check contracts so that releases do not break consumers?

How to justify the choice of CatBoost for tabular time series features and what simple solutions to compare it with?

In PyTorch, what should Dataset do, what should collate_fn do, how do num_workers affect this, and where should .to(device) usually happen?

Multiple threads update individual pixels of the same screen. What can go wrong, and how would you design synchronization?

You pull out visually similar fashion products, but there are a lot of almost identical things on the final list. How do you rank the list to maintain relevance and add variety?

How would you split data for validating a fraud model? What leakage risks would you check?

Review a simple PyTorch multiclass image-classification training loop. What bugs and design issues would you look for?

Design a multimodal fashion compatibility recommender: candidate generation, embeddings, outfit labels, triplet mining, reranking and typical failure modes.

What are the risks of applying reinforcement learning to trading and maintaining market liquidity?

Multiple clients send pixel updates to a central server over the internet. What transport/protocol would you use and what tradeoffs matter?

What techniques can compress or adapt a large model, and how would you detect and reduce catastrophic forgetting during fine-tuning?

How should an agent-generated table be stored so users can sort, filter, and export it without asking the model to regenerate the result?

The user created a presentation, a PDF and a spreadsheet, and then asked to “sort it out.” How do you know what the team is about?

How to test and roll out prompt changes in a LLM product?

What improvements would you add after the baseline real-estate search works: user context, visual embeddings, VLMs, quality models or richer item representations?

How to design a API method that calls an unstable external service or a long pipeline and should gracefully survive failures?

How would you handle new users and new posts in a social-feed recommender with text and image content?

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

A product manager wants an ML model, but there is little or no labeled data. How would you approach the problem?

Explain scaled dot-product attention, why Transformers need positional embeddings, how BPE tokenization works, and what LoRA changes during fine-tuning.

Design a similar-items recommender for 1M items when the current collaborative model fails on cold-start items and misses semantic similarity.