Alice is a complex, high-load service based on large language models. Every year, Alice becomes smarter and helps users solve increasingly complex tasks. Currently, we are using large language models (LLMs) for this more and more often.

Each model is good at a specific task, but training a single universal model that will perfectly solve all tasks is very difficult. Therefore, we teach models to work with each other, turning them into agents. Each agent specializes in something of its own: one searches for information on the internet, another controls a browser, a third works with files or applications.

Alice consists of many components and runs on a large family of platforms. We want to create a new unified homogeneous runtime that will meet modern requirements and be able to support both long agent tasks and requests requiring an instant response.

What tasks await you

Designing Alice's architecture You will design Alice's architecture and develop a new runtime so that Alice can work uniformly with both long tasks lasting tens of minutes and fast queries requiring hundreds of milliseconds without unnecessary overhead. Alice must remain reliable, efficient, and scalable, handling tens of thousands of RPS. It works with diverse types of input data: files, images, voice, text. The runtime will allow improving the system as a whole and individual components (models, tools, agents), and trajectories will be written in unified terms. The architecture must be transparent and understandable.

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

We expect that you

• Know how to develop high-load services
• Know the intricacies of asynchronous communication in multi-component systems
• Understand how agents work
• Are ready to write in C++

Will be a plus

• Have worked with Temporal or another orchestration system
• Are interested in LLMs and have an idea of how they work