Направления, темы и вопросы из записей интервью. Фильтры ниже сохраняются в ссылке.
72 questions from real interviews
There is a robot cashier: the user places a tray of food, the system takes a photo and in a couple of seconds should recognize the dishes and generate a receipt. How to design a ML system?
When is RLHF or DPO needed for a multimodal model, and how to collect preference data for such training?
When is RLHF or DPO needed for a multimodal model, and how to collect preference data for such training?
Design a system that receives an operator-client call recording and automatically fills a booking table with whether the client accepted, the branch, date and one-hour slot.
How to extract structured operations from text PDFs of different banks without sending sensitive data to external services?
How would you handle geography in free-form real-estate queries and keep retrieval fast for millions of listings and high QPS?
A large neural network RecSys-model gives good offline quality, but it must be kept in random. Optimize what?
The project needs to replicate the poorly documented legacy protocol. How to approach research and implementation if part of the behavior has to be reconstructed from traffic and the old system?
There is a pipeline of exchange events: price, exchange timestamp, local timestamp and two primary/secondary delivery channels. How do you know if everything is okay with your data?
In the ranker you need to add new features of the product, user and request. What is offline and what is online?
You have a categorical feature such as port_id. Compare one-hot encoding with historical target aggregates for tree models, and explain the leakage risks.
The stream processor calculates the CTR across windows and drops in the middle of the hour. How to recover without losing events and double counting?
How would you design an LLM-agent loop that checks a task output using tools such as file reading, web access or document inspection?
If the old product used filters rather than free-form text, how would you train a query parser or query encoder before real text-query logs exist?
How to talk about replication, min.insync.replicas, acknowledgements and CAP trade-off at the Kafka/queue level if you need to avoid losing messages in case of failures?
The system is deployed in two data centers, the target SLA is above 99.95. What architectural solutions help not to drop the entire product if one node or service fails?
Imagine two video services with billions of videos: one to quickly find safe snippets inside mostly insecure content, the other to find insecure snippets inside mostly secure content. There is almost no markup, resources and time are scarce. How would you structure the process?
In terms of events, the secondary feed arrives faster than the primary. How to characterize these cases and find the cause?
A video analytics product watches kitchen staff and must check whether people follow location-specific safety protocols. The system needs kitchen rules, time of day and staff context. How would you design the approach?
Users can paste either raw text or a URL. Design the ML pipeline for summarizing the content, including routing, chunking and cost controls.
You need to design a recommendation system or ML platform from scratch. How to choose architecture, data, candidate generation and ranking?
What backend components are needed for a LLM product with tools, cache and long tasks?
A speech product collects user audio. How would you filter and route audio snippets for ASR/TTS training data without poisoning the dataset?
Several annotators transcribe the same low-resource-language audio and produce inconsistent text, numbers and addresses. How do you clean or aggregate the labels?