AI Engineer; m​/f​/d

What this role is about

AI has arrived in the German Mittelstand. The distance between a convincing demo and a system a department trusts every single day is still large - and it gets closed by work that almost nobody does properly: defining what "good" concretely means, building the system toward it, measuring whether it hits, and iterating until it holds.

You don't start from zero. You usually get a real starting point, often a golden data set from the department of a German Mittelstand company: real cases, each with the result an experienced expert would deliver. Your job is to build a system that reaches that result reliably, to bring it into people's working day through an interface they can actually use, and to be able to prove that it works.

You typically work on two to three customer projects in parallel.

• Derive a defensible definition of "correct" from reference data, and build an eval that measures it
• Design and build the AI system itself: model and provider choice, context and prompt design, structured outputs, tool calling, agent/workflow logic, retrieval where it earns its place - and deterministic logic where an LLM isn't needed.
• Do error analysis and iterate deliberately instead of guessing - and demonstrate the improvement
• Build guardrails, escalation paths and monitoring, so the system doesn't fail loudly and confidently
• Build the production interface that goes with it: review screens, approvals, human-in-the-loop - in Next.js and Type Script, not a throwaway prototype
• Roll the solution out and keep it running + improve it as AI advances

• Capture and pass on your state in a structured way: what is built, what was measured, what is still open - in writing, without anyone having to ask
• Explain complex technical matters so that the project lead and colleagues without an AI background can decide on a sound basis
• Raise questions, blockers and wrong assumptions early instead of collecting them until the next meeting
• Occasionally demonstrate or explain a result at the customer yourself

Your standard is not "it's built". Your standard is: it is demonstrably good, it gets used, and it creates real value for the customer.

Three boundaries and all three exist to protect your build time.

• You don't own the customer. The relationship, the workshops and the conversations with management sit with the project lead. You see customers occasionally - when a result needs to be demonstrated or explained - not weekly.
• You don't run the project - and it does get run. We work with structured project management: scope and deadlines are defined up front and actively managed, so you know what you're delivering and by when, and your work stays plannable instead of absorbing whatever shifted this week. Very little ticket boards, no worklogs, no status lists on your side. What we do need: that your state is legible to everyone else at any time, without anyone having to ask.
• You don't work inside our customers' infrastructure - no SAP customizing, no system administration, no legacy integration as your core work.

• Weeks 1-3:
  You work inside two running projects, get to know our stack and our eval practice, and take over your first AI ise case of your own.
• Weeks 4-8:
  You own a complete use case - from the reference data set through the evals to the production interface - and you record results and measurements so that the project lead can represent them without you.
• After three months:
  You build your use cases independently, decide on approach and scope, and are the person on the project team who translates technical topics for everyone else.

• You have built an AI solution that real users used in production - not just a prototype, not just a concept.
• You know how to engineer production LLM systems: context and prompt design, model selection, structured outputs, tool calling, failure handling, latency/cost trade-offs - and you know when a deterministic component beats an LLM.
• You work eval-driven: you define up front what a good result is, you measure systematically, and you don't ship on gut feeling. Whether that runs on Langfuse, Promptfoo, Arize Phoenix or your own spreadsheet is up to you; that you do it at all is not.
• You have built a system in which AI agents plan and execute tasks: tool calling, structured outputs, human-in-the-loop.
• You build production frontend:
  React and Next.js (App Router, Server…