Wolfram Helwig
Physicist. I make AI dependable enough to do real work.
Language models are the most capable source of ideas and drafts anyone has built. They are also a source of disorder: fluent, tireless, and indifferent to whether they are right. Turning that into something a company can rely on is an entropy problem, and entropy is a physicist’s subject.
The connection is literal. A language model picks each word from a probability distribution. The dial that controls how freely it picks is called temperature. That is not a figure of speech: the formula is the Boltzmann distribution from thermodynamics.
So a model on its own behaves like a hot gas of plausible sentences. Fluent, and free to go anywhere.
language modelp(word) ∝ exp(score / T)thermodynamicsp(state) ∝ exp(−E / kT)
The structure around the model is called a harness. It gives that gas a shape: the right sources, limits on what the model may touch, memory, and checks that say whether the result is right.
Building harnesses is my work. Alpentropy, alp plus entropy, is my company.
What a harness is made of
- Access
- The model reads the actual repository, document or mailbox, not its memory of similar ones.LLM Harness, MCP servers
- Limits
- Reading, writing, running tests, Git and sending mail are separate rights. The narrowest one that works is the one it gets.LLM Harness, Gmail server
- Memory
- A long task survives a restart. What was learned in one session is there in the next.mnemo, resumable tasks
- Checks
- Scripts where a check can be exact, model judges where it needs judgement, and never the author judging its own work.Comic Workshop, review agents
The projects behind it
- LLM Harnesssince 10/2026
- Connects language models to authorised repositories, documentation and tickets, for source-backed analysis and controlled engineering work. Separate permissions for reading, writing, running tests and Git. Tasks can be resumed, retries are bounded, failures are recorded.
- mnemo05/2026 to 07/2026
- A memory engine for AI coding agents: short-, mid- and long-term memory across sessions, with nightly consolidation runs that leave an auditable report.github.com/StochasticEntropy/mnemo
- MCP servers2026
- A local Gmail server that gives agents several mailboxes and attachments. It writes drafts and does not send unless told so explicitly. And a small lab server that makes each part of the Model Context Protocol visible.
- Comic Workshop09/2026 to 10/2026
- A production line where Claude agents and an image model made a 114-page illustrated book through to the print PDF. Most of the work went into checking: scripts where a check can be exact, model judges where it needs judgement.
- Developer toolingsince 2025
- robot-companion, a VS Code extension for Robot Framework, and runcomposer, a platform for orchestrating large automated test suites. Both are open under the MIT licence.github.com/StochasticEntropy
Background
I did my PhD in physics at the University of Toronto, on entangled quantum states. After that came game theory and reinforcement learning, and since 2017 software, cloud and machine-learning work in industry: automotive, e-commerce, smart mobility, insurance.
What physics left me with is a habit: measure before you believe. With language models that habit is useful, because they are fluent whether or not they are right.
What you can hire me for
- Harnesses and agent systems around language models
- LLM applications with tool integration over MCP
- Evaluation: deterministic checks and LLM judges
- Machine learning, data and cloud engineering
- Training and workshops
Remote-first. On site in Luzern, Zug and Zürich. German or English.
Contact
Everything stated on this page is in my CV or in the linked repositories.