Running Claude Code on a local model, measured
A measured benchmark on an M5 Max, plus the three criteria that decide whether local inference is worth it for Claude Code.
Blog
What building with AI agents actually looks like: tools that earned their place, mistakes that cost time, and decisions I'd make differently today.
A measured benchmark on an M5 Max, plus the three criteria that decide whether local inference is worth it for Claude Code.
A write path into a foreign system has six stages. The costly mistake: the word gate names two categorically different things.
A second session discarded the first one's uncommitted work. Why none of it came back, and three precautions that would have helped.
I write the longer reports from my projects on my personal site. These are the ones that go with a course.
Six screenshots sent on the go become four tasks. How Sven, Navigator and session-orchestrator work, and what I still decide.
Related course: Multi-Agent Orchestration in Practice
A RAM warning fired on 84 percent of all session starts with zero OOM events. The fix was never the threshold, it was measuring the right quantity.
Related course: The Digital Caretaker
· in German only
Überlappende Konfidenzintervalle heißen nicht „gleich gut“. Gepaart gegen eine feste Marge verglichen, hält der billigere Ersatz auf keiner von sechs Aufgaben.
Related course: Verify AI output reliably
An agent loop needs more than repetition. I explain how checks, bounded corrections and clear stopping conditions turn repeated attempts into useful work.
Related course: Loop Engineering
Graph engineering connects work, data and checks across AI agents. A concrete example shows where parallel work helps and what makes a result ready to use.
Related course: Multi-Agent Orchestration in Practice
A render call could take 46 seconds inside a webhook with a 30 second budget. The second finding turned out to be worse than the first one.
Related course: From Laptop to Live System