I architect systems that automate expert knowledge work.
And I design them to be deterministic, traceable and explainable: the reasoning lives in formal ontologies and rules, language models are confined to language, and people make the decisions.
Axibra, a Methodology Operating System for strategy consulting.
Axibra delivers the UNITE Body-of-Knowledge, more than two hundred strategy frameworks, methodology modules and pattern libraries over a formal ontology, as a runtime for consultants. I own its architecture as Founding CTO, and I do the product development, the ontology engineering and the Go development that goes with it.
Graph queries over the ontology, a rule-based term index, an algorithmic DAG orderer and byte-deterministic catalog vectors. The guarantee does not depend on a model returning the same answer twice.
Every AI call lands in a trace store. Every model output passes a consultant review gate before it is used.
Because the reasoning is a graph and a set of rules, every result can be walked back to the frameworks and relations that produced it.
fusion. Agentic development as an engineering discipline, not a demo.
fusion is the harness I built to develop software with coordinated AI agents and human developers in one team. It is the method behind Axibra, and it is separate from it. The articles trace how it came about.
- Specialised rolesRequirements design, planning, implementation, review, state audit, analysis, policy curation, orchestration. Each role has one prompt and one job.
- Shared conventionsEvery agent and every developer works to the same rules, recorded in an append-only event log that is never rewritten.
- Human gatesNothing lands without a decision by a person. Agents prepare; people rule.
- Team coordination via gitfusion splits work so that dependencies between work items are respected and merge conflicts are reduced, for humans and agents alike.
- Two adaptersOne for Claude Code. One for model APIs: Anthropic, OpenAI, and locally running models. No vendor lock-in, and local models for environments that require them.
Three decades of software, AI and semantics.
I have built software since the early days of the web, from startups to government projects. A postgraduate degree in computational linguistics and a PhD in the humanities are the reason semantics and ontology engineering run through all of it as one continuous thread.
I am a single father. I care about human values and stand for democracy, human dignity, freedom and equality.
Lead developer and technical project lead at Fraunhofer IAIS. Metadata of 300 million cultural-heritage objects integrated over CIDOC CRM. The platform is still in operation and became the technical basis of the ARD Mediathek.
An event-centric model of my own design, as the ontology that harmonises globally distributed databases into one data warehouse.
UEOF and UIF, the formal ontology under a Methodology Operating System. Same approach, now as the substrate of an AI product.
Seven pieces on building with AI, written while building fusion.
They read best in order: how Claude Code works and how to extend it, why agile breaks when AI writes the code, the model that replaces it, its foundation, an essay on what the tools do to their makers, and a report from the implementation.
- Understanding Claude Code
- Extending Claude Code
- Beyond Agile: Project Management for AI-Driven Development
- Foundation: Konzeptionelle Grundlage für AI-gestützte Projektarbeit
- Fusion: A Process Model for AI-Driven Software Development
- The Maker's Paradox
- Implementing Fusion: A Process Layer for AI-Driven Development
Tooling I use myself, published as tenzoki.
fusionThe agentic development harness.flightOne calm agent for non-technical users: documents, discussion, precise written output.scoutCited multi-source web research, one report per run.stilwerkStylometry, authorship attribution, style transformation.callscopeCompiler-grounded change impact for Rust workspaces.
gooidcOIDC and JWT utilities for Keycloak and Entra ID.godastKey-value, graph, content-addressable storage and search on BadgerDB.gblobsDeduplicated, optionally encrypted blob store.cuecastStateless BPMN-subset process engine.krkTwo-pane file manager with editor for macOS, Norton-Commander style, with git view.