How to Train Your Domain Model - Tobias Goeschel - DDD Europe 2025

Why It Matters

If LLMs can reliably encode domain rules, software development cycles could shrink dramatically, reshaping how enterprises build and maintain mission‑critical applications.

Key Takeaways

• LLM agents can execute domain logic tasks
• Experiments test LLMs against deterministic business rules
• Fine‑tuning with DDD models improves reasoning accuracy
• GenAI may reduce code generation overhead
• Results show promising but limited reliability

Pulse Analysis

The latest generation of large language models (LLMs) arrives equipped with agentic capabilities and built‑in guardrails, promising higher accuracy and the ability to perform autonomous tasks. For enterprises, these features open the door to more than just code‑completion; they enable systems that can reason, plan, and act within complex business environments. As generative AI moves from experimental labs into production, architects are evaluating whether these models can handle the deterministic logic that underpins mission‑critical applications. This shift challenges traditional development pipelines, prompting a reassessment of how software is designed, tested, and maintained.

Tobias Goeschel’s DDD Europe 2025 presentation frames this debate through the lens of Domain‑Driven Design. He demonstrates a series of experiments where LLMs are fed DDD artifacts—UML diagrams, bounded‑context descriptions, and ubiquitous language glossaries—to fine‑tune their internal representations. The results indicate that models can translate textual domain specifications into executable logic, but they still struggle with edge‑case invariants and strict transactional guarantees. By comparing model‑generated outcomes against deterministic rule engines, the study quantifies accuracy gaps and highlights the importance of guardrails and human oversight in the loop.

The implications for software vendors are twofold. First, successful integration of LLM‑driven domain reasoning could dramatically shorten development cycles, allowing teams to prototype business capabilities directly from high‑level models rather than hand‑coding every rule. Second, the current reliability ceiling suggests a hybrid approach, where AI‑generated components are validated by traditional testing frameworks before deployment. As the industry refines fine‑tuning techniques and expands the corpus of DDD‑aligned data, we may see a gradual migration toward AI‑augmented architecture, reshaping roles from pure coders to AI‑orchestrators.

Original Description

Domain-Driven Design Europe 2025 - Organised by Aardling (https://aardling.eu/)

https://dddeurope.com

https://newsletter.dddeurope.com/

https://be.linkedin.com/company/domain-driven-design-europe

https://bsky.app/profile/dddeu.bsky.social

https://mastodon.social/@[e-mail]

The newest generation of Large Language Models come with exciting features, such as agents and guardrails, which promise great improvements in accuracy, and enable GenAI applications to execute a variety of tasks.

Instead of using models to write code, can we use them to replace code? How well are they suited to reason about complex domains? Can they keep up with traditional (deterministic) domain logic? Can they be fine tuned using DDD artifacts, such as graphical or textual representations of the domain?

This talk shows a series of tests and experiments that may bring us closer to embracing GenAI as a general way of designing business applications.

About Tobias Goeschel:

Senior Solutions Architect at AWS

Tobias started his career as a freelance web developer in the late 90s and has since worked on hundreds of projects of varying sizes and lengths - from a single person to multiple teams, from a few days to several years - and in many different roles: Consultant, crafter, coach, and... well, architect. He is a strong advocate of diversity and inclusion in the tech industry, and an active member of the European Software Crafters and Domain Driven Design communities.