Generator line

The tools have changed: UML tool, in-house generators, MDA frameworks, Xtext, AI agents. The task stayed the same: generate code from a specification and check the result.

The stages are described neutrally; I do not name clients or projects. The periods come from the profile model.

1. 1997–2000

   Select Software Tools (vendor of UML tools): automated data preparation from the UML tool SELECT Enterprise into Word, Lotus Notes, Cold Fusion; SQL generators

   Today's reading

   Documents and SQL were generated from the model in the UML tool. The first generation of the same task.

2. 2004

   Code generator for test drivers in a load-testing project

   Today's reading

   Generating tests instead of writing them. The same applies to agents: the machine produces draft tests, the human adds the edge cases.

3. 2008–2009

   Model-driven migration of an ERP system (MDA)

   Today's reading

   The model as the source of truth, maintained across the entire life cycle of a product.

4. 2010–2014

   CRM portal (telecommunications): unified SQL generation from Excel specifications, coding conventions, machine-checked quality rules

   Today's reading

   Code is generated from the specification, and a machine checks the rules. Today these are the guard rails an agent is measured against.

5. 2014–2015

   B2B interface (telecommunications): migration of a generator framework to Xtext, modelling workbench, generators for WSDL/XSD, verification by AST and text comparison

   Today's reading

   Checking generated output against a reference, via the syntax tree and not just the text. Every agent that generates code raises the same task.

6. 2016–2018

   Information system of a federal authority: in-house tool for semi-automated UI migration, automated migration tests

   Today's reading

   Semi-automated means: the tool does the routine part, the human the rest. With agents, this division of labour is on the agenda again.

7. 2019

   Preliminary study on data exchange between authorities: context analysis with a uniform DSL, XSD → Ecore, canonical event format

   Today's reading

   First the context boundaries, then a shared language for the analysis. Without both, an agent has nothing to hold on to.

8. 2021–present

   Systems for plausibility checking and planning at a federal authority: model-driven product line, specification language for rule-based planning, architecture documentation

   Today's reading

   The language of the business domain becomes the grammar of the system. A specification that a parser reads leaves less room for misunderstanding, for humans and agents alike.

9. since 2026

   bewirb.es: three DSLs, code generation, M2M/M2T, built in dialogue with an agent

   Today's reading

   This website is generated from this project. How it works is described under This site.

10. since 2026

    Specification-driven development with AI agents (AI Unified Process), extending a legacy system of a federal authority with a new business area: vertical slice as a requirements elicitation tool, corrections only in specification and guidelines, never in generated code; rule set formalised from a sister service; agent environment opencode and Claude Code

    Today's reading

    The specification is the leading artefact, the code its reproducible derivation. What is new is the generator, not the method.