Банк вопросов из реальных собеседований

What is data drift, how would you detect it in production, and what can you do when it starts hurting an ML model?

Какие логи, метрики и алерты нужны после запуска рекомендательной модели?

Tell us what classes of models are in recommendation systems and where they are usually used.

You have a multi-GPU server and want to host one or more open-source LLMs. What software stack and design choices would you use?

The cart changes during the session. How should the service handle added and removed products, cache invalidation, and fallback responses?

Where are the main costs and errors in creating rare scenario simulations and how to use them safely in a search engine model?

How do online and offline recommender designs differ when model latency, feature cost and freshness matter?

How to organize an online heavy recommendation model and meet a set delay budget?

A delivery-pricing model needs a fresh measure of courier load. Which events and states should it use, and how would you make the feature safe for online serving?

How would you profile and optimize an LLM inference pipeline across batching, latency, throughput, memory, and cost?

A neural network inference pipeline is too slow. What optimizations would you consider before changing the model architecture?

The call-processing pipeline works but inference is too expensive. What would you optimize first?

What main architecture families are used for generative models, and where are they commonly applied?

What should a ML engineer do to bring the model to a production service: interface, model file, Docker, monitoring and updates?

What does it mean to take an ML model from training to production, and which pieces should an ML engineer be able to own?

Your punctuation model has good validation metrics, but users complain that some outputs are strange. What could be happening and how would you debug it?

What are the differences between recall@K, precision@K, coverage and NDCG in evaluating recommendations? How do they behave when they change K?

How would you evaluate the full search pipeline and its individual components offline and online?

We need to build a system where the advertiser looks at the CTR of campaigns. Given 200 billion impressions per day and a CTR of about 1%. How to start system design with numbers?

The photo shows the organization's sign. How to build a pipeline that extracts text and uses it in a product?

A 72B-parameter LLM is served on an A100 80GB. Estimate whether FP16 fits and explain what quantization changes.

The catalog and photos are constantly changing. How to retrain the model and update the visual search ANN index?

After launching a feed recommender, how do you decide when and how to retrain the models?

How would you define requirements for a platform that trains large models on temporarily idle GPUs in multiple data centers?