Developer in Advertising Ranking

To select the best banner for a query, we use multi-stage ranking, various algorithms, and services for ML model inference: generative models, neural networks on GPU/CPU, BERT, decision trees. At each stage, numerous ML models are calculated to narrow down the banner funnel. Then, at the auction stage, multi-banner blocks are formed and compared with each other to determine the winner and send it for rendering to the user.

We roll out changes every six months that increase advertising revenue, our systems handle around a million RPS, and our inference services process approximately 100 million banners per second.

What tasks await you

Development in the engine — adding new stages to auctions, implementing block filling algorithms We are actively reworking auction mechanics: implementing new model architectures, changing approaches to auction amnesty, researching block filling algorithms, adding features, and optimizing performance.

Support for new CPM prediction models — feature preparation, using microservices for inference We have projects related to ML model inference services. Models in different services are calculated in parallel to reduce critical execution time. Each of the services is optimized to increase throughput.

More about backend at Yandex — in the channel Yandex for Backend

We expect that you

• Are proficient in C++
• Know concurrency
• Are familiar with ML fundamentals
• Have worked with microservices
• Are familiar with profiling

It will be a plus

• Have worked with GPU
• Have gained MLOps experience