Банк вопросов из реальных собеседований

Команда говорит, что переходит от single-node vector search к distributed vector retrieval system. Какие вопросы и trade-off стоит обсудить?

The inDrive super app has several verticals: taxi, delivery, intercity travel, freight transportation and couriers. How to recommend an active user of one vertical to try another? Understand metrics, data, model, offline evaluation and A/B test.

Design an automatic system that checks whether a human/agent task result is good enough before delivery to a customer. How do you frame the ML problem?

How would you train a feed ranker from likes, comments, shares, bookmarks and non-clicked impressions?

Design real-estate search without explicit filters: the user enters a free-form text query. How would you define metrics, listing features, query/user context, and the first retrieval/ranking setup?

In the marketplace there was a carousel or tag "Offer of the day" with model selection of ads, and conversions increased. How to separate the model effect from the new interface effect?

How to train users to match: which architecture, loss function and target to choose if users are recommended other users?

OCR API processes one document in 2 seconds, while processing takes 1 CPU core and 3.5-4 GB RAM. The server has 20 cores and 64 GB RAM. How to calculate safe throughput?

For a FastAPI-backed LLM product, when would you use Postgres, ClickHouse and Redis?

What commands can you use to check files, logs, processes, resources, port, and system service when an application crashes?

How do merge and rebase combine the history of branches, how do their effects differ, and when is it safe to apply each method?

The panel should show CTR by minute, five-minute and hourly windows. How to design API and aggregate storage?

On review you see a test that only checks HTTP 200. What's wrong with it and how can you make the check useful?

How to solve a cold start for a new user in the recommendation feed? When to use popularity, user-based, item-based, and content-based approaches?

What is important when running a CV model on an edge device or VR headset?

How do dropout and BatchNorm behave in train vs inference, and what can go wrong when fine-tuning a network with small batches or multi-GPU training?

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

Kafka topic has 10 partitions and 100 tasks: 90 tasks of 90 ms each and 10 tasks of 1 s each. Tasks are distributed evenly across partitions, and within the partition the order is consistent. How to estimate the best/worst completion time for 1, 10 and 20 consumers?

How does LightFM help cold start and when do bandits appear in recommendations?

OCR recognized the text on the sign. How to understand which organization it corresponds to and when the result can be published?

The team is developing a multimodal model for search and product like neuro-responses. Which use cases should be chosen and how to prioritize?

What does a “representative photo” mean for an organization in a geo-product and how to select such a photo automatically?

What if a business change was recorded in the database and the event in Kafka did not go?

The product may have several photos: general appearance, details, different angles. How to get a sustainable view of the product for search by photo?