I take AI products from conception to launch.
At Cybus I lead the agentic-AI initiative: shipped Connectware GPT 0→1, our LLM assistant now in public beta, and built our MCP server 0→1, an API surface for AI agents working with millions of live data points from major German industrial players.
Beyond shipping, I own what makes AI products trustworthy: guardrails, governance, and output quality. On a platform that runs factories, "mostly correct" isn't good enough.
Products I've shipped
AI Product Manager
Leading the agentic-AI initiative on Connectware, an industrial data platform that connects factory equipment to IT systems for major German industrial players.
- Shipped Connectware GPT 0→1: the platform's LLM assistant, now in public beta.
- Built the Cybus MCP server 0→1: an API surface that lets AI agents work with millions of live industrial data points.
- Own AI trust end to end: guardrails, governance, and output quality for every AI feature.
Product Manager, HealthTech
Product manager at wellabe, a Munich HealthTech company, working on digital health products where user trust and regulatory care shape every decision.
Four things I believe about building AI products
Ship the zero-to-one, then make it trustworthy.
A first version proves the idea. Guardrails, governance and output quality are what turn it into a product people rely on. I own both halves.
"Mostly correct" is a demo, not a product.
On a platform that runs factories, an LLM answer that is right 90% of the time is a liability. I treat evaluation as part of the product, with the same rigour as any other production system.
Agents need an API surface, not a chat box.
The useful AI products I have built expose capabilities (tools, data, permissions) that agents can compose. The conversation is the last mile, not the architecture.
Math first, hype second.
Applied mathematics and computer science are where I started. It shows up in how I frame problems, read model behaviour, and decide what to measure.
Side projects and experiments
Write-ups of side projects and experiments are in progress. Until then, the best evidence of how I build is in my work at Cybus.
Read the case studiesNotes on building AI products
I'm writing about what it takes to ship LLM features that people can trust in production: evaluation, guardrails, and agent tooling. Until it lands here, I share shorter notes on LinkedIn.
Follow on LinkedIn ↗