AI engineer in Frankfurt Rhine-Main

What an AI engineer does

An AI engineer builds software on models that already exist. Chip Huyen defines AI engineering as “the process of building applications with readily available foundation models” (AI Engineering, O’Reilly, 2025). A lab such as Anthropic or OpenAI trains the model, and the AI engineer connects it to a company’s systems and decides what it may do there.

AI engineer, ML engineer or software engineer

Job ads mix these titles, and each one means different work. The table shows what each role builds and which part of it I cover.

Role	What the role builds	My part
AI engineer	Software on existing models: agents, tools, MCP servers and the checks around them	My daily work
ML engineer	Its own models, trained on the company’s data and run in production	Not my field; I do not train models
Data scientist	Analyses and statistical models that answer business questions	Not my field
Software engineer	Code for a product, with or without a model	I write code as well, from APIs to web portals
Forward deployed engineer	Solutions inside one customer’s organization. Gergely Orosz describes engineers who “embed with customers” and “co-develop solutions” (The Pragmatic Engineer, August 12, 2025)	Close to my client work, where I build inside one client’s systems
AI architect	The design of how models, data and systems fit together	Part of my work as IT architect in Frankfurt

What I build with Claude and Codex

Most of my AI engineering today happens in one client engagement, on an information platform built with AI agents. These are the parts I set up there:

• Parallel sessions. Several sessions of Claude and Codex and the human team work in one repository at the same time, each session in its own git worktree.
• Hooks with tests. A hook blocks a dangerous command before it runs, and a test matrix proves that it lets valid work through.
• A gateway for Claude on a phone. The control plane of the agent team speaks MCP over stdio only. I put an OAuth gateway, built with FastMCP, in front of it, so Claude on a phone reads and writes the team’s issues.
• A team of agents with roles. On Paperclip a chief agent hands each task to one agent of its team, and each code agent opens a merge request under its own name.
• A gate before code changes. An agent changes code only after it has queried a knowledge graph of the repository for the part it touches.

The rules behind this setup are on AI agents under the same rules as humans.

Who I am

I am Michael Wutzke. I have built software for more than twenty years, from online shops to insurance portals on AWS, and today I build it with AI agents such as Claude and Codex. I am also interested in open-source AI models that a company runs on its own servers. Earlier I was Head of Decentralized Finance and Node Operations, then Chief Information Officer, at the Frankfurt-based company Blocksize Capital. I hold the AWS Solutions Architect Associate certificate and take part in Claude Certified Architect Foundations, the first official certification program of Anthropic. Details: Teaching and certifications.