The course covers the fundamentals and application of foundation models in civil engineering and architecture: how transformer-based models work, how they are trained, fine-tuned and prompted, and how LLM-based agent systems are built. In the accompanying exercises and an individual project, students develop their own agentic software engineering project and use AI tools both for code generation and for code review, while critically examining the results themselves.

New in the winter semester 2026/27. The language of the course is English.

Module Description 

The full module description is available on TUMonline.

Organisation

Classes take place each Thursday, 9:45-11:15, in the Seminarraum Architekturinformatik (0501.04.170C). The first class is on Thursday, 15.10.2026. The course is worth 3 ECTS. Due to the high individual supervision effort, the number of participants is limited.

Assessment consists of a project report including source code (50%, approx. 8-12 pages), submitted at the end of the lecture period, and a written e-test of approx. 30 minutes (50%).

Recommended prior knowledge: basics of machine learning and deep learning, programming skills in Python, and confident use of version control (Git), command line and development environment.

Contents

- Principles of language processing
- Transformer architecture and the functioning of modern foundation models
- Pre-training, fine-tuning and alignment (RLHF, PEFT, LoRA)
- Prompt engineering and inference-time strategies
- Advanced foundation model architectures: reasoning models, multimodal LLMs, linear attention
- Agentic software engineering: human-AI collaboration in code creation and code review
- LLM agent systems (e.g. tool use, retrieval-augmented generation (RAG), memory)
- Foundation models for use cases in civil engineering (e.g. automated compliance checking, generative design)

Literature

- Raschka, S. (2024): *Build a Large Language Model (From Scratch).* Manning.
- Huyen, C. (2025): *AI Engineering - Building Applications with Foundation Models.* O'Reilly.

Contact

Dr. Stefan Fuchs - stefan.m.fuchs@tum.de