Experiments
Experiment 1, August 2026
I put six different brains inside the same agent
Six local LLMs, the same tools, the same questions. I measured how differently they behaved.
Read the experiment (opens another site)
On the bench
One agent, two frameworks
The same agent built in LangGraph and in CrewAI, then run on local models. What does each framework cost?
LangGraph, CrewAI, Ollama
Where the tokens go
The same kinds of questions sent through one agent, compared model by model on token usage.
granite4.2, qwen2.5, qwen3.5
The instrument
PCPilot
The local desktop agent behind these experiments. It streams from Ollama and acts through tiered tools, with path guards, deny lists, a permission ledger and an undo journal.
Python, PyQt6, Ollama
About
At Hach in Düsseldorf, I work on the services that connect IoT devices to the cloud through Azure IoT Hub and Service Bus. I've also helped drive AI adoption there, including AI-assisted coding, automated code reviews and workflow automation.
I have more than ten years of full-stack experience across .NET, React and Azure, and I've worked with AI and ML since 2020.