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

Bremen, Berlin

Full-time

Permanent employee

Your tasks

Most people who train models for a living only ever get a slice of the run: the data, or the infra, or one ablation. Here you'd get the whole thing, start to finish.
We train models for agentic work in European companies and public administration. That means long tasks, real tools, and users with very little patience for "mostly right". The recipe is reinforcement learning on a big, varied pile of synthetic tasks, running on hundreds of GPUs across EuroHPC and commercial compute.
We're looking for the person who makes that training their own. Someone who makes the models more capable and more reliable, and who gets a little grumpy when a GPU-hour is wasted.

What you'll work on

Post-training, end to end, from SFT warm starts to big on-policy RL runs on agentic tasks. It's yours, dull stretches included.
A better recipe: algorithms, curricula, data mixtures, KL and entropy control, credit assignment over long horizons. You ablate properly and keep only what survives.
Models that are cheap to run as well as good. Fewer tokens and tool calls per solved task, lower latency, a smaller bill for whoever deploys them.
Finding out why training is lying to you. A reward curve that looks great for the wrong reason. Entropy that collapses. A rollout worker that quietly drops trajectories.
The stack underneath: rollout and inference engines, async training, and runs and evals that someone else can reproduce, whether they sit on EuroHPC's Slurm machines or on commercial compute.
Shipping. Some models go out in the open (the first release is planned for Q1 2027), others go to customers who depend on them.

Your profile

You've trained or fine-tuned LLMs with your own hands, ideally with RL (the PPO/GRPO family, RL with verifiable rewards, RLHF).
PyTorch is second nature, and you've done distributed training for real (FSDP, DeepSpeed, Megatron or similar) on Slurm or Kubernetes clusters.
You're disciplined about experiments: a hypothesis, a baseline, enough seeds, and plots you'd defend in front of a skeptic.
You don't much care which side of the research/engineering line a problem falls on, and you like simple things that work.
You can say clearly what a run showed. That includes the runs that showed nothing.
You live in the EU and are allowed to work here. English is all you need; German is a bonus.
Nice to have, and truly optional: time with inference engines like vLLM or SGLang; RL frameworks (verl, OpenRLHF, TRL, or the one you wrote because the others annoyed you); multi-node debugging (NCCL, interconnects, hardware that dies mid-run); GPU performance work; papers or open-source contributions.

Why us?

The recipe works, and it's nowhere near finished. We think that's the best time to join: late enough that you're not starting from an empty repo, early enough that your judgment still changes how we train. You won't be on your own with it, either. The team has been training models for years, through LeoLM, OpenEuroLLM and SOOFI, so you won't be the only one here who has stared at a loss spike at 3 a.m. And we're small enough that you'll recognize your own work in the model.

What we offer

Work from anywhere in the EU, ideally Germany.
Real compute: hundreds of GPUs across EuroHPC and commercial compute.
Whatever AI tools make your work easier, without a procurement saga.
Your name on public work. Our first open-source release is planned for Q1 2027.
Founders you can just message, and a research team worth learning from.
An office if you want one: Bremen, right on the water (yes, there's a boat, and free drinks and smoothies), or Berlin.

How to apply

Upload your CV through the form and tell us, in a few lines, about one thing you built that you're proud of. An environment, a training run, a judge, a bug that took you three days to corner. Links to GitHub, papers, a blog post or a messy notebook are all welcome. Please don't write us a polished cover letter. We'd rather read the notebook.
We're hiring for three RL roles at the same time: environments and data, post-training, and alignment and reward modeling. If you're somewhere between two of them, apply to the closer one and say so. We'll figure it out together.
We care a lot more about what you've built and how you think than about how your CV reads. If you don't tick every box, apply anyway. Hardly anyone does, and different perspectives make our models and our team better.

About us

At ellamind, we develop language models and the methods that make them better. Our team trains its own models, builds RL environments, datasets and benchmarks, and researches how to reliably measure and systematically improve the capabilities of LLMs. We are known for our leading German-language open-source models. We actively share what we learn, including through the open-source community DiscoResearch and as part of the OpenEuroLLM program.

As a cash-flow-positive startup, we work with well-known partners such as Deutsche Telekom, the German Federal Government and leading healthcare providers. Our headquarters is in Bremen. We offer competitive pay, equity options and a short line to the founders.

Your CV isn't the only thing that counts. We are looking for curious people who want to help shape the next generation of European AI models. Diversity matters to us, because different perspectives make us better. Please apply even if you don't (yet) meet every requirement.