ML Engineer for the Online Learning of Generative Personalization Group

We create recommendation systems that help millions of users find exactly what they need. Our team works at the intersection of recommendation systems, deep learning, and infrastructure, reliably implementing cutting-edge neural network approaches in high-load services handling thousands of RPS.

What we offer: * High technical and research culture * Convenient tools for prototyping and conducting experiments * Hundreds of top-tier GPUs and petabytes of logs for model training and inference * Varied projects and development of expert experience in the field of recommendations across different domains within the team * The opportunity to implement E2E solutions that directly impact users

We also share the results of our research and speak at conferences (RecSys, KDD).

Our articles: * Scaling Recommender Transformers to One Billion Parameters * Correcting the LogQ Correction: Revisiting Sampled Softmax for Large-Scale Retrieval * Yambda-5B: A Large-Scale Multi-modal Dataset for Ranking And Retrieval

What tasks await you

Speeding up RL fine-tuning algorithms for recommendation models We know how to improve recommendation quality by using Reinforcement Learning (GRPO, DPO) to quickly adapt models based on fresh user feedback. You will be creating and optimizing a runtime pipeline that will allow retraining and rolling out updated neural network models in tens of minutes. The key focus is on accelerating RL cycle iterations and efficient use of computational resources.

Building and implementing distributed RT-processing For training and serving modern recommendation models that work with extensive user profiles, reliable and efficient sample delivery is necessary. You will expand existing and create new systems for Yandex services that ensure the storage, transfer, and processing of such data with guaranteed high availability and uninterrupted operation of RL pipelines.

Implementing the Feature Store concept to improve ML pipelines How to ensure identical feature processing for neural network models in real-time inference and fine-tuning modes? We are developing a framework, already implemented in the Music and Pictures domains, that guarantees such consistency. You will expand its functionality and scope of application, paying attention to the specifics of working with data for reinforcement learning.

We expect you to

• Be proficient in Python and C++, including having confident optimization skills
• Understand the principles of distributed data processing and MapReduce paradigms
• Know the basics of neural network operation
• Be able to understand someone else's code and propose architecture improvements

It will be a plus

• Have experience in developing and debugging multi-threaded or high-load systems