Research Scientist / Engineer, Alignment & Reward Modeling (all genders)

Your tasks

Picture an agent working through cases at a public authority. It's fast, and it's mostly right. Every so often it also invents a file number, skips a check it was supposed to run, or quietly bends its instructions to finish sooner. For the people we build for, that agent is worse than no agent at all. They're right about that.
So much of our work comes down to one unglamorous question: how do you teach a model to be reliable, and how would you know if you hadn't? That's this job. You'd build the reward models and judges that decide what "good" means during training. You'd train models to be honest about what they did and didn't do. And you'd build the evaluations that catch them when they aren't.
You'd be inside our RL training, not on a separate team reviewing it afterwards. What you build goes straight into the reward.

What you'll work on

• Reward models and LLM judges for work a unit test can't grade: rubric-based, generative, process-level. Then checking them against careful human judgment, because a judge nobody has checked is just another model with opinions.
• Turning the behaviors that matter in regulated work into training signal. Following policy and instructions when it's inconvenient. Asking when something is ambiguous. Stopping instead of guessing. Reporting what a tool actually returned. Knowing how sure to be.
• Hunting reward hacking and specification gaming, and studying how our judges fail so we can fix them.
• Alignment experiments inside RL post-training (preference data, synthetic critiques, principle-based feedback), with an honest account of what they cost in capability and efficiency.
• Safety and reliability evals that still mean something outside the lab. We do evaluation for enterprises and public-sector clients, so there are real cases to draw on.

Your profile

• You've trained or evaluated LLMs, and got your hands dirty in at least one of these: reward modeling, RLHF or preference learning, LLM-as-judge, alignment or safety evaluation.
• Your Python and PyTorch are strong enough to take an experiment from idea to result on your own.
• You're suspicious of metrics. When a number looks good, your first thought is how it could be wrong.
• You can take a fuzzy complaint ("it's overconfident on edge cases") and turn it into something we can measure and train against.
• You write clearly. An alignment result nobody believes hasn't achieved much.
• You live in the EU and are allowed to work here. English is all you need; German is a bonus.
• Nice to have, none of it required: interpretability or scalable-oversight work, publications on reward modeling or evaluation, time spent in a regulated field, some familiarity with the EU AI Act, German.

Why us?

Here, alignment is the reason anyone can use the product at all, so your work ends up in every model we ship. You wouldn't be starting cold, either. The team has spent years on evaluation, from benchmark tasks for Terminal-Bench 4.0 to safety and performance evaluations of models for the German federal government. There's enough compute to test an idea properly instead of arguing about it. And our first open-source release in Q1 2027 is a chance to show, in public, what a reliable European agent model looks like.

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