Agentic AI workflows

    AI agents that do the work, with people in charge

    I design and build agentic AI workflows for small and mid-sized teams: agents that handle a real process inside your tools, log every step and hand consequential decisions to a person. Because I also do EU AI Act engineering, the oversight is part of the design, not added later.

    Workflow scoping

    We pick one process that costs your team time today and work out whether an agent should do it at all. I map the steps, the systems involved, where a mistake would be expensive and where a person must stay in charge.

    You get a written design: what the agent does, what it never does alone, the tools it connects to and one metric for success.

    Agent development

    I build the workflow in Python with LangGraph, LangChain or CrewAI, on AWS Bedrock or self-hosted models, and connect it to the tools your team already uses through their APIs.

    You get a working workflow running in your environment, tested against real examples from your process.

    Oversight built in from day one

    Every step is logged, the agent's permissions are limited to what the task needs, and anything consequential waits for a person to approve it. If the EU AI Act applies, the same components cover Article 50 disclosure and the Articles 12–14 duties.

    You get an agent you can explain, audit and stop, not a black box.

    Handover

    The code is yours. I document how the workflow works, how to change its prompts and limits, and how to read its logs, then walk your team through it.

    You get the code, documentation and a handover session so your team can run and extend it.

    Work you can check

    Two of my agentic projects are published as open-source Python packages:

    I've also published research on what the EU AI Act means for agents: Article 50 compliance for agentic AI: nine unresolved engineering questions.

    For an AI workflow designed so people make every decision, see Then and now: keeping people in charge of hiring (an illustrative redesign, not client work).

    For how I scope an AI project with a client, read the in-house support chatbot case study.

    Who I work with

    Small and mid-sized companies and scale-ups in Germany and the EU that want to automate a specific process with AI but don't have an in-house team to build it. Engagements are remote. I'm one engineer, so I take on a small number at a time.

    Questions

    What kind of workflows do you build?
    Multi-step AI workflows that do real work inside your existing tools: triaging and answering support requests, processing documents, preparing drafts for review, or automating developer tasks such as pull request review. Each one is scoped to a single process with a measurable result.
    Which tech stacks do you work with?
    Python, with LangGraph, LangChain or CrewAI for orchestration, on AWS (Lambda, Bedrock) or with self-hosted open-source models. I connect agents to your systems through their APIs. Other stacks are possible; ask.
    How do you keep an agent under control?
    Every workflow is built with logging of each step, clear limits on what the agent may do on its own, and a human approval point before anything consequential happens. If the workflow falls under the EU AI Act, the same components cover the Article 50 disclosure and, where relevant, the Articles 12–14 duties.
    Do we own the code?
    Yes. The workflow runs in your environment, and you get the code, documentation and a handover session so your team can maintain and extend it.
    How are projects invoiced?
    Through SMartDE eG, a cooperative for self-employed professionals in Germany. Its details are in the Impressum.

    Tell me about the process

    Describe the task your team wants off its plate and the tools involved. I'll tell you honestly whether an agent is the right answer.

    Get in touch