R2D: From Requirements to Design: Leveraging LLMs for Augmented Software Design

Name of participant: Tobias Eisenreich

Project’s name: From Requirements to Design: Leveraging LLMs for Augmented Software Design

Project description:

Software architecture design remains a largely manual, fragmented process. While established methods such as Domain-Driven Design (DDD) exist, they are complex and tedious. At the same time, emerging AI tools often lack methodological depth, leaving a gap between practical usability and engineering rigor.

My research project addresses this gap by developing an architecture design process with AI support. The goal is to transform natural-language requirements into well-structured, evaluated software architectures with known properties, using large language models (LLMs) as collaborative “junior architects” rather than autonomous decision-makers. With the Technical University of Munich (TUM) for scientific guidance and the industry partner DATEV, I am well equipped for both scientific rigor and practical relevance.

At its core, the project integrates three established techniques into a unified pipeline. Starting from textual requirements, I use Domain-Driven Design (DDD) to model the problem domain and establish a shared understanding of core concepts. Building on this, Attribute-Driven Design (ADD) systematically derives the software architecture based on quality attributes and design constraints. Finally, the resulting architecture is evaluated using the Architecture Tradeoff Analysis Method (ATAM), which assesses trade-offs, risks, and quality scenarios. All steps are supported or automated using generative AI. As expected, our first experiments show that expertise of a human architect stays crucial to avoid hallucinations and error buildup during these steps.

The main innovation lies in combining these established methods with LLMs’ flexibility within a single, coherent workflow. In contrast, traditional model-driven approaches force users into a rigid process with very specific steps, while very flexible systems like AI chatbots cannot provide the guidance or integration that software developers need. This enables both creative exploration and structured reasoning. By basing the tool on proven processes, it can guide human users, preventing more junior users from getting lost in the vast amount of possibilities.

Technically, the project follows a modular approach: Individual prototypes for each step are developed and evaluated, then seamlessly integrated into a unified platform. An industrial case study with DATEV follows to ensure practical relevance and real-world applicability.

The expected impact is twofold. From a research perspective, the project provides a foundation for systematically studying AI-assisted architecture design. From an industry perspective, it lowers the barrier to applying advanced methods like DDD, ADD, and ATAM, reducing effort and improving the quality and consistency of architectural decisions. In the long term, the platform can serve as an extensible basis for future AI-native software engineering tools. Furthermore, it could be a first step towards further automation, potentially enabling vibe coding to develop advanced architectures with little guidance.

Software Campus Partner: TU München and DATEV

Implementation period: 01.01.2026 – 31.12.2027