We are looking for an ML engineer with strong practical experience with LLMs: fine-tuning, post-training, evaluation, and inference.

Important: this is not a classic ML Engineer / Data Scientist position. We need a candidate who has independently trained and adapted LLMs and understands the full cycle — from data and training to evaluation and deployment.

Tasks:

• Development and improvement of LLM post-training pipelines: SFT, preference optimization, Offline / Online RL;
• Improvement of instruction following and building a benchmarking / evaluation system;
• Training LoRA / QLoRA adapters for various domains and tasks;
• Working with reward models / verifiers and reward functions for RL;
• Development of RAG systems: retrieval, reranking, context construction, and evaluation;
• Training and integration of tool use / function calling / structured outputs;
• Optimization of model inference and serving using vLLM / SGLang;

Mandatory requirements:

• Practical experience in LLM fine-tuning / post-training;
• Experience with SFT and preference optimization: DPO / IPO / ORPO / SimPO or equivalents;
• Experience or good practical understanding of Online RL for LLMs: GRPO / GSPO or similar methods;
• Understanding of reward models, preference data, and LLM evaluation;
• Confident command of Python, PyTorch, Hugging Face Transformers;
• Experience with TRL and/or verl;
• Experience with LoRA / QLoRA / PEFT;
• Experience with inference / serving via vLLM and/or SGLang;
• Experience in developing RAG, tool use, or structured generation;
• Understanding of modern LLM architecture and specifics of their training and inference;
• Experience with GPUs, distributed training, and Linux / Docker / Git;
• Ability to independently set up experiments, select metrics, and analyze results.

Will be a plus:

• Continual / domain pre-training;
• Reasoning RL;
• Multilingual LLMs or code generation;
• FSDP / DeepSpeed / Megatron-LM;
• Quantization and inference optimization.