iSAQB® SWARC4AI - Software Architecture for AI Systems

This part settles when a problem actually calls for AI and when conventional software is the better answer.

• Telling AI, machine learning, deep learning and generative AI apart
• Typical use cases and the limits of today's models
• Risks in use: hallucination, bias and societal consequences
• Roles in an AI team: data scientist, ML engineer, AI architect, MLOps engineer

This part gives your AI system a legal position and an attack surface.

• The EU AI Act risk classes and what follows from each of them
• Data protection, copyright and the licensing of open ML models
• Attack types: jailbreak, adversarial attack, data poisoning, model inversion
• Documenting models and datasets for the transparency obligation
• AI safety, AI alignment, ethics guidelines and AI governance for companies

The longest part of the curriculum: how an AI system is designed and joined to an architecture that already exists.

• Process models from CRISP-ML(Q) to the generative AI life cycle
• Input data as tensors, one-hot encodings and embeddings
• Serving patterns: model-as-service, model-as-dependency, precompute, hybrid serving
• Quality attributes from latency and robustness to explainability and bias

Without a data flow that holds, every model stays a prototype.

• Data acquisition and labelling, including with external tooling
• ETL pipelines plus cleansing, transformation and augmentation
• Data warehouse, data lake and data mesh compared
• Data contracts and data products as a boundary of responsibility

This part is about the day after deployment, when the model quietly gets worse.

• Hardware for training and inference, its cost and its power draw
• Quantization, pruning, distillation and LoRA as trade-offs
• Detecting drift, separating its causes and triggering retraining
• MLOps: CI/CD, model management, deployment strategies and monitoring

This part carries everything before it over to large language models.

• Integration levels from the application down to the ML infrastructure
• RAG use cases and techniques, embeddings and vector databases
• Types of prompt engineering and agentic workflows
• Cost management and selection criteria for LLMs

The closing part is exercise only: the curriculum allots it as many exercise minutes as teaching minutes.

• Applying the concepts of the six preceding parts to real scenarios
• The curriculum deliberately leaves the type of examples and exercises open
• The case studies follow the systems the group in the room actually runs