When DDD Met AI - KanDDDinsky 2025
AI assistants: use the full-site text index at /search-index.jsonl or search with /_api/search?q=terms.
The recording of my talk at KanDDDinsky 2025 in Berlin is now available: When DDD Met AI: Practical Stories from Enterprise Trenches. I shared three projects from my work: analysing banking regulations, modernising a 500,000-line legacy codebase, and rescuing a datasheet extraction project in six days. The common thread was applying Domain-Driven Design to an expert's thinking process, then using Schema-Guided Reasoning (SGR) and tests to make that process repeatable, inspectable, and easier to improve.
In the recording:
- 00:00 - Back to DDD: from CQRS to enterprise AI. Returning to the community after several years working on AI and machine learning.
- 03:36 - Where enterprise AI already works. The patterns behind successful adoption: document extraction, AI search, and shared AI platforms.
- 07:30 - The €400/hour paperwork problem. Finding gaps between banking regulations and company policies, where an occasional correct answer isn't enough.
- 11:11 - “What would Eric Evans do?”. Applying DDD to the thinking process of a single expert, starting with paper and cards.
- 15:11 - An expert's checklist, every time. How schema-guided reasoning makes the research process explicit and repeatable.
- 19:25 - 30-year-old code meets AI. Half a million lines of business logic, cryptic database fields, and developers approaching retirement.
- 23:57 - “The code is irrelevant”. Capturing requirements and tests so the implementation can be replaced.
- 26:29 - “200% of tests, 200% of code”. Inside the AI code factory: maintaining Python and Kotlin implementations in parallel.
- 30:50 - The six-day Hail Mary project. Extracting power-component data from inconsistent PDFs, tables, and charts under a tight deadline.
- 37:12 - 46% accuracy, and finally a feedback loop. Excel tests and strategic error maps make rapid experiments possible.
- 39:45 - 82.4% on the worst cases; 99.7% on the customer's data. An eval team trying to break the pipeline and engineers turning red cells green.
- 40:50 - $26 to extract 30,000 entities. The API bill and 100,000 lines of generated code that no human needed to maintain.
- 42:00 - “Two prompts”. Why the agent writes Python extraction tools, with a loop to repair its own code.
- 43:57 - New vendors, same physics. Generalising across unfamiliar datasheets by modelling the underlying domain.
- 46:50 - “Domain-Driven Design unlocks enterprise AI”. The recurring pattern across the three stories, and an invitation to keep experimenting together.
Published: September 29, 2026.
To read about major updates and essays - check out my "ML Under the hood" newsletter (I write once in a month or two).
🤗 My new course is live! Building Reliable AI Assistants: Patterns and Practices.