∇Services
Services
Mandates in software engineering, data science, AI systems and training. Remote-first, with time on site in Luzern, Zug and Zürich for kickoffs, workshops and milestones.
|ψ⟩: many agents in parallel, one measured result
AI harnesses & agents
LLM applications with real tool access, separate permissions and honest evaluation. Agent teams that work in parallel and get checked independently.
Language models are fluent, and fluency is not evidence. Agent systems built here cite their sources, say “unknown” when they should, run with the narrowest permissions that work, and are verified by checks that did not write the code.
Scope
- LLM applications with tool integration over MCP
- Agentic development workflows: parallel agents in isolated working copies, independent review
- Evaluation with deterministic checks and LLM judges
- Local-first AI tooling where data must not leave the machine
Track record
- LLM Harness: source-backed software analysis with separate read, write, test and Git permissions
- Comic Workshop: agentic production line for a 114-page book with mixed deterministic and agentic QA
- mnemo: a memory engine for AI coding agents, with nightly consolidation runs and auditable reports
Toolbox: Claude Code · Codex · MCP · LLM judges · Python · TypeScript · Git worktrees
∂system/∂t: change, under control
Software engineering
Backends, APIs, cloud infrastructure and the pipelines that ship them. Python first, test-driven where it matters.
Most software fails slowly, one undocumented assumption at a time. The answer is small, explicit contracts, an automated path to production, and code a team can keep changing without fear.
Scope
- Green-field architecture and backend development in Python
- Serverless and containerised services on AWS and Azure
- Infrastructure as code, CI/CD and automated test pipelines
- Test automation at scale, including Robot Framework tooling
Track record
- Green-field Python architecture for a large-scale test automation project (insurance)
- Serverless backend on AWS Lambda processing streaming data (smart mobility)
- API-driven data management on AWS with DynamoDB and S3 (e-commerce)
Toolbox: Python · FastAPI · pydantic · SQLAlchemy · TypeScript · React · AWS · Azure · Docker · Kubernetes · Terraform · Jenkins · GitLab
signal over noise
Data science & machine learning
From exploratory notebook to model in production: classical ML, deep learning and reinforcement learning, on AWS or Azure.
A physicist’s instinct for data: model the system before you fit it, quantify the uncertainty, and distrust any result that looks too good. Then put the model where it earns its keep, in production, monitored.
Scope
- Exploratory analysis and modelling with a real modelling component
- Deep learning for vision and language, including transformer models
- Reinforcement learning, simulation and optimisation
- ML pipelines and deployment on Azure ML, Databricks or AWS
Track record
- Containerised BERT-based NLP model for text classification in a serverless architecture
- Deep reinforcement learning and game-theoretic methods for imperfect-information scenarios
- Microsoft Certified: Azure Data Scientist, Data Engineer and AI Engineer Associate
Toolbox: PyTorch · TensorFlow · JAX · scikit-learn · pandas · NumPy · Spark · Databricks · Azure ML · SageMaker
∫ knowledge dt, integrated into your team
Training & workshops
Hands-on courses in machine learning, deep learning and agentic development, for teams who want to do it themselves.
The best outcome of a workshop is a team that no longer needs the trainer. Courses are built around notebooks and real exercises, from the first regression to transformers, and from the first prompt to a reviewed agent workflow.
Scope
- Machine learning fundamentals with Python, NumPy, pandas and scikit-learn
- Deep learning: CNNs, RNNs and Transformers
- Agentic software development with Claude Code and Codex
- Test automation with Robot Framework
Track record
- Interactive course on classical ML and a deep-learning module with TensorFlow (automotive)
- Tutorials in theoretical physics and for undergraduate students at two universities
Toolbox: Python · scikit-learn · TensorFlow · PyTorch · Jupyter · Claude Code · Codex · Robot Framework
∇How a mandate runs
- t₀
Measure
Understand the system as it is: code, data, people, constraints. No solution before the problem has a shape.
- t₁
Model
Write down the contracts, the unknowns and what “done” means, in a form a check can verify.
- t₂
Build
Small increments, often in parallel with supervised coding agents in isolated working copies.
- t₃
Verify
Independent checks against the evidence before anything counts as done. Then hand over, documented.