NLP Developer for Alice

Skills: NLP. Specializations: ML Developer.

We are a team of ML researchers and engineers within Alice. We are building Russia's first autonomous computer-use agent that can interact with computer environments like a human: click, scroll, fill out forms, switch between applications, and autonomously solve user tasks.

We already have a working prototype in production, an agent platform for millions of users, GPU clusters, environment simulators, and training pipelines using GRPO-like approaches. We need someone to take the agent's RL training to the next level.

What tasks await you

Design and launch experiments for training an agent to interact with computer environments You will be responsible for devising and implementing new RL pipelines — from formulating reward functions to the architecture of training on multi-step trajectories with tool calls. You will work with environment simulators and real trajectories: building a data flywheel from data collection to model improvement. You will need to train and fine-tune large models, and optimize inference for production: FP8, distillation, parallelism.

Improve GUI grounding and visual screen understanding (VLM components of the agent) You will participate in designing the agent's architecture: memory, planning, tool orchestration. You will need to pay attention to safety and truthfulness aspects, bringing R&D results to production on a platform serving millions of users.

More about ML at Yandex — in the Yandex for ML channel

We expect you to have

• Deep practical experience with DL and RL: trained models using GRPO, GSPO, DAPO, Dr.GRPO, understand reward design nuances, know why training breaks and how to fix it
• Trained large models — engaged in fine-tuning LLMs and VLMs not by following tutorials, but in real-world tasks
• Possess a research mindset: able to find and critically evaluate papers, extract applicable ideas, and quickly test hypotheses
• Work with a task from experiment setup to production results
• Proficient in Python, PyTorch

It will be a plus if you have

• Worked with VLMs: trained, fine-tuned, or performed inference with them
• Worked with multi-turn RL on trajectories of interaction with tools and environments
• Have publications at top conferences: NeurIPS, ICML, ICLR, ACL, etc.
• Worked with computer use, web agents, GUI grounding