Research Engineer / Scientist, Agentic RL Post-Training (all genders)

ellamind • Berlin, KY

4d • Remote

About The Position

At ellamind, the Research Engineer / Scientist will own end-to-end post-training for models designed to handle agentic work in European companies and public administration. The work spans supervised fine-tuning (SFT) warm starts through large on-policy reinforcement-learning runs on agentic tasks, using varied synthetic tasks and compute across EuroHPC and commercial infrastructure.

The role involves improving model capability and reliability, developing and evaluating training recipes, and diagnosing training failures and misleading metrics. It also includes making models more efficient to deploy, building reproducible rollout, inference, training, and evaluation systems, and shipping models for open release and customer use. The first open-source release is planned for Q1 2027.

The position can be performed from anywhere in the EU, ideally Germany. Optional office locations are Bremen and Berlin. The team works across research and engineering, and values disciplined experimentation, clear communication of results, and practical solutions.

Requirements

• Hands-on experience training or fine-tuning LLMs, ideally with reinforcement learning such as PPO/GRPO, RL with verifiable rewards, or RLHF.
• Strong practical experience with PyTorch.
• Experience with distributed training, such as FSDP, DeepSpeed, Megatron, or similar, on Slurm or Kubernetes clusters.
• A disciplined approach to experiments, including hypotheses, baselines, sufficient seeds, and defensible plots.
• Comfort working across research and engineering problems, with a preference for simple, effective solutions.
• Ability to clearly communicate what experimental runs showed, including inconclusive results.
• Residence in the EU and authorization to work there.
• English proficiency; German is not required.

Nice To Haves

• Experience with inference engines such as vLLM or SGLang.
• Experience with RL frameworks such as verl, OpenRLHF, or TRL, or a custom RL framework.
• Experience debugging multi-node systems, including NCCL, interconnects, and hardware failures during runs.
• GPU performance optimization experience.
• Research papers or open-source contributions.
• German language skills.

Responsibilities

• Own end-to-end post-training, from SFT warm starts to large on-policy RL runs on agentic tasks.
• Improve training recipes, including algorithms, curricula, data mixtures, KL and entropy control, and long-horizon credit assignment.
• Design and conduct ablations, retaining changes supported by experimental results.
• Improve model efficiency by reducing tokens and tool calls per solved task, lowering latency, and reducing deployment costs.
• Diagnose training issues, misleading reward curves, entropy collapse, and rollout workers that drop trajectories.
• Work on rollout and inference engines, asynchronous training, and reproducible runs and evaluations across EuroHPC Slurm systems and commercial compute.
• Ship models for open release and for customers.

Benefits

• Work from anywhere in the EU.
• Access to hundreds of GPUs across EuroHPC and commercial compute.
• Access to AI tools that make the work easier.
• Opportunity to be credited on public work.
• Direct access to the founders and the opportunity to learn from the research team.
• Optional offices in Bremen and Berlin; the Bremen office has free drinks and smoothies.
• Competitive pay.
• Equity options.

What This Job Offers

• Job Type: Full-time
• Education Level: No Education Listed