Interview Practice Questions

Describe the embedding model training in more detail: architecture, data, and metrics.

What augmentations did you use when training the vision model?

А как модели распознать физически, что картинки одинаковые?

From what you said, have you worked on both video computer vision tasks and more conventional computer vision tasks?

What tasks and projects are related to EVE (virtual reality)?

Is there a model for first-person recognition?

Is the task about video pixelization?

Are first-person videos spherical videos?

How to implement 2D to 3D video conversion using depth estimation and projections?

How to implement AI Editor system for automatic video processing by moderators?

How to automatically cut VR video into short segments for TikTok?

How to implement Outpainting for expanding video to spherical format?

How to implement Six Degrees of Freedom for creating volumetric video from regular cameras?

What else do algorithms detect? What errors in images?

How do you work with NSFW content in the product?

Can additional information be extracted from text and images?

Design a system that will collect and store features for a machine learning model. Ensure that the system can handle 10 million records per day and must provide access to features with a latency of no more than 100 ms. The team consists of 3 members.

Build a system for monitoring and alerting on data drift. The system should handle real-time data with a volume of 1 million events per day and must send alerts in case of drift detection with a delay of no more than 5 minutes. The budget is $5000 per month.

Design a system for experiment tracking using MLflow. The system should support up to 1000 concurrent experiments with a total database size of up to 500 GB and must provide access to results with a latency of no more than 10 seconds.

Build an A/B testing system capable of handling 5 million requests per day with two user groups. The system must collect and analyze user interaction data with a delay of no more than 1 hour and provide the ability to dynamically adjust tests.