Senior AI Evals Engineer (AI Quality & Safety) - Libra-Legal AI Assistant (m/w/d)

About Us

There are hundreds of thousands of lawyers across Europe, and at Libra we're

transforming how they work. Our AI platform combines deep legal reasoning

with cutting-edge generative technology, fundamentally changing how lawyers

research, draft, and deliver legal work.

We operate with the speed and ownership of a startup, while being backed by

the scale and stability of an established leader. Libra is an independent

business unit within the Legal & Regulatory division of Wolters Kluwer, a leading

global provider of information, software, and services for professionals. For 180

years, Wolters Kluwer has supported and simplified the work of experts and

organizations through innovative solutions, relying today on more than 20,000

colleagues worldwide to bring that vision to life.

About The Role

As an Evals Engineer (m/w/d) at Libra, you'll be the first dedicated hire on a new

team that owns how we measure quality: the datasets, rubrics, judges and

harnesses that decide whether an AI feature is good enough to ship, and the

automated optimization that runs against them.

Optimizing a system against a target is newly automatable: LLM optimizers like

GEPA now propose and test the candidates themselves. Nobody has to invent

the experiments any more, and the system will improve in whatever direction

the evals point, whether or not that's where you meant to go. That makes

defining what good means the highest-leverage work we do, and it has to be

done task by task, from scratch, for a legal AI platform at European scale.

You'll work in a lean, agile environment with AI Engineers, Legal Engineers and

Product, based at the vibrant Merantix AI Campus in Berlin, surrounded by a

community of AI innovators.

What you'll do

• Build and own the eval platform: datasets, judges, harnesses, regression gates, cost and quality in one view, on Python, FastAPI and Langfuse. Make it self-service, so AI Engineers can evaluate and tune their own features without going through you, and eval-driven development becomes the most effective path rather than a tax.

• Map the quality landscape: good means something different for research, drafting, summarisation and retrieval, and again per jurisdiction. Work out what a defensible measure looks like for each, going first on the ones nobody has evaluated before and then making them repeatable without you.

• Design the rubrics and set the standard for LLM-as-judge: turn Legal Engineers' and subject-matter experts' judgment into version-controlled criteria, keep judges recalibrated as models and jurisdictions change, and make authoring cheap enough to do at volume on privileged material.

• Deep-dive results and traces until you can say why something failed, then find the lever that moves it and automate the fix.

• Unhobble the optimizer: instrument the app so prompts, hyperparameters and harness architecture become levers a search can safely pull, then run automated optimization (GEPA, DSPy) over them against a fitness function you trust, widening that surface as you go.

• Make cheaper models win: treat quality per euro as a first-class metric, instrumented per call, and find the configuration where a smaller model matches or beats an expensive one.

• Own guardrails: ungrounded advice, invented or misattributed citations, jurisdiction and language leakage, prompt injection from ingested documents. Design them, red-team them, prove they hold.

What you'll bring

Education

• Bachelor's degree or equivalent in a relevant technical field (e.g. Computer Science, Software Engineering, Statistics, Data Science); advanced degree is a plus.

Experience

• Minimum 5 years in software engineering,at least 1 building LLM-powered products in production.

• StrongPython: FastAPI, modern tooling, and the data stack (pandas, numpy, notebooks), because much of this job is analysis.

• Hands-on experience designing evaluations: datasets, rubrics, LLM-as-judge, benchmarking, human labelling.

• Solid security and data-privacy practice. You'll handle traces, documents and datasets derived from privileged legal material.

• AI coding agents (Claude Code, Codex, Cursor) in your daily workflow.

• Genuine interest in the legal domain.

• Bonus: automated prompt or pipeline optimization (GEPA, DSPy or similar), and a broader data science toolkit, e.g. embedding clustering to check dataset coverage.

Skills

• A strong engineer who hasn't hand-written code in months. The architecture is yours, the typing isn't. You ship more working software than you ever did alone, reject code that runs but is shaped wrong, and leave less rework and cognitive debt behind you.

• You think in systems and expect them to be gamed. The app, the evals, the optimizer and the people using them are one loop. Once evals gate releases, everything optimizes toward them, so some judgments stay human.

• Hard to fool by a single number. "Could these all be within variance?" comes before any ranking, and a judge's reliability before you trust its grades. You report bounds, and retire a result that doesn't hold up, including your own.

• You get expertise out of people who have no time to give you any, and build tools they choose to use. An eval nobody runs is worth nothing.

• Runs on macro-management. You ask the sharp questions up front, then come back with options, their trade-offs and the assumptions behind each, and help pick. Good to think out loud with. Entrepreneurial, accountable, pragmatic.

• Excellent communication in English.

Our Interview Practices

To maintain a fair and genuine hiring process, we kindly ask that all candidates participate in interviews without the assistance of AI tools or external prompts. Our interview process is designed to assess your individual skills, experiences, and communication style. We value authenticity and want to ensure we’re getting to know you-not a digital assistant. To help maintain this integrity, we ask to remove virtual backgrounds and include in-person interviews in our hiring process. Please note that use of AI-generated responses or third-party support during interviews will be grounds for disqualification from the recruitment process.

Applicants may be required to appear onsite at a Wolters Kluwer office as part of the recruitment process.