Work with me
Agents that hold up in production.
I help teams take an agent workflow from a promising demo to daily use, and measure what it costs, how good the output is, and where it breaks.
Where agent projects stall.
Costs nobody planned for
The demo ran on the biggest model. At real volume, retries and long contexts multiply the bill.
Measure cost per accepted output and route each step to the cheapest model that is good enough.
Output nobody trusts
When nobody knows how often the agent is wrong, a person ends up checking everything.
Make failure measurable, add checks where errors cost money, and send only real judgment calls to a person.
Models that change every month
Prices drop, models get retired, and a better option appears weeks after you have committed.
Keep the model choice swappable and re-test it against your own cases, not a leaderboard.
Where to start
An agent reliability review.
One workflow you run today or plan to run, measured end to end. Fixed scope, agreed before we start.
- 01
Map
One workflow, its inputs, its quality bar, and what a failure costs you.
- 02
Run and measure
Run it many times on your real cases. Record cost, output quality, retries and where it breaks.
- 03
Fix and decide
The fixes worth making, in order, and a clear call: ship it, change it, or stop.
What you get
- Cost per accepted output, not per token
- How often a person had to step in, and why
- The failure modes, ranked by what they cost
- Which model each step actually needs
If the review says the workflow is worth building or fixing, I can build it. If it says it is not, that is the answer you get.
A working example: this site.
The LLM and AI video comparisons run on the same approach: agents gather prices and benchmark results from many sources, reconcile what disagrees, and nothing is published until it passes checks.
From source to published number
Last run Nov 4, 2026 · replay
Agents do the legwork. Checks block what doesn't reconcile, thin evidence is marked provisional, and judgment calls go to a person.
Background
Independent AI engineer with a PhD in physics. I have built production systems for German enterprises in health insurance, banking, medical associations and product data, mostly turning messy documents and data into something a business can rely on.
Not every problem needs an agent, or AI at all. If a simpler approach does the job, I will say so. More on the about page.
Have an agent that needs to work every day?
A few sentences about the workflow and where it hurts are enough to start.