Curriculum for

Certified Professional for
Software Architecture (CPSA)^®
Advanced Level:
Module
SWARC4AI

Software Architecture for AI Systems

Participants will know how artificial intelligence is defined and how machine learning, data science, deep learning and generative AI are categorized within it.

Participants will know how to specify typical general use cases for AI. This includes use cases for e.g. image recognition & generation, language processing, prediction, personalization and anomaly detection.

Participants will be familiar with the possible applications in various industries, e.g. marketing, medicine, robotics and content creation. They will also gain an overview of the possible uses of AI in end-user applications. These include, for example, voice assistants (chatbots) and recommendation systems (recommender engines).

Participants can identify risks that arise when using AI. These can be the following risks, for example:

Participants understand the differences between AI systems and traditional software:

Participants can decide and explain whether and why a problem should be solved using either AI or classic software development.

The participants know typical roles and their tasks in these contexts. The roles include in particular

Data Scientist, Data Analyst, Data Engineer, Machine Learning Engineer, MLOps Engineer, AI Architect, Data Architect, Business Intelligence (BI) Developer, Data Governance Specialist and ML Researcher.

Participants will be able to identify and prioritize AI use cases.

Participants understand the strengths and limitations of AI and are familiar with the so-called "Jagged Technological Frontier".

This phenomenon helps to understand why companies can initially experience a drop in productivity when implementing AI, but this can be followed by productivity gains as it develops further.

Participants are familiar with data protection laws such as the GDPR and know how they affect the collection, processing and storage of data by AI systems.

The participants understand the objectives and regulations of the EU AI Act and know what influence this has on the development and use of AI systems. They also understand the requirements of the EU AI Act (Trustworthy AI) and what influence these requirements have on the development process and the architecture of the software system. In particular, they know the influence on some of the following aspects:

The participants know the classification of AI systems according to the EU AI Act risk levels (prohibited, high-risk, limited-risk, low-risk) and know which regulatory requirements apply in each case.

The participants understand the copyright issues for AI-generated content, know the effects on certain software license models and can devise potential solutions for these issues.

The participants gain an overview of different types and degrees of openness of free ML models. This concerns, for example, the disclosure of data and model parameters. In addition, they know different types of licenses of free ML models and their impact on the AI system.

The participants understand the basic statements of the EU AI Act (in particular transparency obligations) and know strategies for compliance and possible challenges.

Participants know how to effectively document models and data sets to ensure traceability and transparency.

Participants will be familiar with possible security pitfalls and typical types of attack on ML models. This applies, for example:

Participants know how to develop and apply strategies to minimize AI risks. In particular, the participants know the following strategies:

Participants will be familiar with various options for protecting against attacks (AI security) and, in particular, they will be familiar with options for integrating security standards into the architecture and can take these into account for the design of the system.

The participants understand the basic problems of AI safety and know the various facets of it. In particular, participants will be familiar with specific problems such as "AI model risks by poisoning" and "bias".

Participants are aware of the ethical problems that AI systems can bring with them and know approaches and possibilities for resolving these problems. This includes, for example, AI alignment (and its limits) and the creation of their own AI guidelines.

The participants review important ethics guidelines such as the "EU Ethics Guidelines for trustworthy AI" and the "Google AI Ethics Guidelines".

The participants are familiar with the most important documents on AI governance in order to develop the core principles of AI governance and responsible AI for the company. This applies in particular to the following documents:

Participants will gain an insight into the creation of regulatory sandboxes to promote innovation and the possible legal consequences of non-compliance with the provisions of the AI Act.

Participants will know how effective data management ensures the quality and security of data in AI applications.

Participants have an understanding of the life cycle of a machine learning or data science project. In particular, they are familiar with the following phases:

Participants will be familiar with typical process models for the software development of AI systems. These are, for example

Participants know different types of data and typical ML problems as well as use cases for utilizing the data. In addition, the participants have an understanding of various requirements for the data, e.g. the presence of correct labels of various types.

The participants understand the various machine learning problems such as

and know what requirements they have.

Participants know which data is required for AI algorithms and have a good understanding of the need for validation and knowledge of the typical data breakdown into training, validation and test data. In addition, participants have a good understanding of the input data required for various AI algorithms such as

are required.

Participants understand how to deal with challenges such as non-determinism, data quality and concept and model drift.

The participants are familiar with transfer learning and fine-tuning as a way of incorporating the pre-trained basic models into existing use cases.

Participants will know which design patterns exist for AI systems and how to select suitable patterns. This concerns the following design patterns:

The participants know how to define the use cases / tasks of an ML model, such as the classification of images or fraud detection.

Participants understand which inputs and outputs are required for the functioning of an ML system and can specify them.

Participants are familiar with various metrics for measuring the performance of ML models, such as:

and know how to define evaluation criteria for performance assessment.

Participants understand how ML models can be integrated into existing systems and are familiar with the interfaces and integration points.

Participants know how user interfaces should be designed to enable effective interactions with the ML system and optimize the user experience.

Participants understand the importance of key performance indicators such as latency and throughput in AI systems and know how these can be optimized.

Participants understand the importance of scalability for increased data volumes and know how to develop AI systems that can handle increasing data volumes without losing performance.

Participants have an understanding of what robustness means in AI systems and can apply strategies to increase robustness in different application contexts.

Participants understand the concepts of reliability and availability and know how to build AI systems that are stable and constantly available.

The participants know how important it is that AI results are reproducible and testable and know which methods can be used to ensure these properties.

Participants are familiar with the requirements for security, data protection and compliance and know how these are implemented in AI systems.

Participants understand the importance of explainability and interpretability in AI systems and know how to ensure this in order to promote trust and transparency.

The participants know how bias in data and models can be recognized and reduced in order to ensure fairness and equal treatment in AI applications.

Participants are familiar with the concepts of fault tolerance and can explain how AI systems remain functional despite errors or malfunctions.

Participants gain an overview of various methods for acquiring and labeling data.

Participants can name common platforms for publicly accessible data.

The participants know relevant tools for data labeling such as CVAT, Amazon Mechanical Turk.

The participants have an understanding of the design of efficient data pipelines and data architectures, and they understand how to consider data quality for storage solutions and their management. To this end, participants are familiar with architectural patterns for data engineering pipelines and ETL processes.

The participants know strategies for data aggregation, cleansing, transformation, enrichment and augmentation.

Participants will gain an overview of relevant tools for data engineering pipelines such as Apache Spark and Flink.

Participants will be familiar with various options for storing data as well as their advantages and disadvantages. The options include the following technologies:

The participants know the different (hardware) requirements for TPU, GPU or CPU for training and inference, for example.

Participants will be able to name examples of trade-offs between different model architectures with regard to quality characteristics. In particular for scaling, efficiency and storage load, participants should know the trade-offs as well as the advantages and disadvantages of important architectures such as RNNs and transformers.

Participants will be familiar with ways to adjust various quality characteristics such as accuracy, efficiency and storage load of an ML model and to trade them off against each other. In particular, the participants know the following techniques:

Participants gain an understanding of the costs, power consumption and sustainable use of AI (green IT). In particular, participants will know how to develop AI models and systems that work in a resource-efficient manner by using memory, computing power and storage space efficiently.

The participants know the term MLOps for the automation of the life cycle of a data science project and the connection with DevOps.

The participants have an understanding of tracking in model training, parameters, metrics and results.

Participants will gain an overview of approaches for evaluating ML models and AI systems based on them.

Participants are familiar with different types of drift, such as data drift or model drift, as well as possible causes and solutions.

Participants will have an understanding of CI/CD pipelines, model management and deployment strategies for AI models.

The participants are familiar with common platforms for model provision, such as Huggingface Hub.

Participants will be able to name common tools for creating POCs for AI systems, such as Gradio, and understand how they work conceptually.

Participants are familiar with various deployment options for AI models. Participants will understand a selection of the following options:

The participants know the advantages and disadvantages of SaaS and self-hosting and can weigh up the pros and cons.

Participants will gain an overview of well-known SaaS AI solutions, such as Azure OpenAI Services.

Participants will be familiar with various options and standards for embedded deployments of ML models.

The participants understand the need for monitoring, also with regard to AI-specific requirements such as drift tracking. Participants are familiar with relevant metrics such as accuracy, precision, recall and F1 score, MAE, MSE, perplexity, latency, throughput and resource utilization and understand why these are relevant for monitoring.

The participants know example tools for monitoring. This includes both general tools, such as Prometheus & Grafana, as well as ML-specific tools, such as MLflow.

Participants understand the benefits of user feedback for further model training. In addition, participants are familiar with various methods and tools for collecting user feedback, such as choosing between multiple answers and flagging in Gradio.

Participants are familiar with various methods of using feedback for model training, such as RLHF, RLAIF and DPO.

Using a practical example, participants will learn what an MLOps pipeline can look like and what insights it offers into parameters, metrics, etc.

Participants will be able to make build vs. buy decisions for MLOps systems/components.

The participants are familiar with well-known MLOps tools and end-to-end platforms, such as:

Participants are familiar with the various levels of AI integration. These include the following levels:

The participants know some examples of common libraries, interfaces and tools for the integration of AI models.

Participants know how integrate AI systems into IT landscapes at a strategic level. For example, they can use the strategic design of DDD (especially context maps) to determine and document the type and degree of integration of AI systems.

The participants understand the quality characteristics that are particularly relevant for AI systems. This includes the following quality characteristics in particular:

Participants are familiar with common evaluation frameworks such as LangSmith or LangFuse to deal with indeterminacy and errors in AI systems.

Participants will practise and discuss a case study with an imaginary technicality to weigh up integration options for AI into an existing software landscape.

The participants have a basic understanding of generative AI such as LLMs and stable diffusion.

Participants understand how LLMs work and can categorize the associated terminology: Token, Embedding, RNN, Transformer, Attention.

The participants understand known patterns for the use of LLMs. This includes the following patterns:

The participants know typical use cases for RAG such as "Talk to your documents/database/API".

The participants know a selection of common RAG techniques such as:

The participants know different types of prompt engineering, such as

as well as general best practices for prompting.

Participants know what agentic workflows are and are familiar with the terms reflection, tool usage, planning and multi-agent collaboration.

The participants know which design patterns exist for generative AI systems. They know a selection of the following patterns:

The participants know several techniques for evaluating LLM applications. These can be, for example, the following techniques:

The participants are familiar with well-known LLMs such as GPT, Claude, Gemini, Llama, Mistral or Luminous, and the selection criteria for a suitable LLM.

The participants understand the importance of cost management for GenAI applications.

The participants know some examples of common libraries, interfaces and tools in connection with LLM applications such as OpenAI-API or LangChain.

The participants know agentic AI software architectures, AI agent architecture components, types of AI agent architectures.