Generative AI: Theory and Practice (ICT302)

This course equips students with the engineering knowledge and practical skills to design, build, evaluate, and critically assess agentic AI systems powered by large language models (LLMs). Beginning with the foundational concepts of transformer-based LLMs, including context windows, tokenisation, and embeddings; students progress to the core disciplines of agentic engineering: prompt engineering for structured outputs, tool use and function calling, agent loop and harness design, loop design patterns, memory and retrieval-augmented generation (RAG), and multi-agent orchestration. Provider-agnostic design principles are emphasised throughout, enabling students to apply their skills across different LLM platforms. The course addresses reliability through the design of evaluation suites, and safety through the analysis of prompt injection in agentic contexts. Ethical considerations, data privacy, and responsible deployment are examined in relation to systems students have built, rather than as abstract principles. Students also engage critically with AI-assisted development practices, documenting and reflecting on the use of AI coding tools throughout the course.

• The LLM landscape: models, capabilities, context windows, and engineering context
• Transformers for practitioners: tokenisation, attention, and embeddings
• Prompt engineering as a discipline: structured outputs, templates, and AI-assisted development
• Tool use and function calling: schemas, dispatch, and the Model Context Protocol (MCP)
• Dialogue-based LLM applications: The agent loop and harness architecture
• Loop design patterns
• Memory and context management
• Multi-agent orchestration and provider-agnostic design
• Evaluation and reliability engineering for agentic systems
• Safety in agentic systems
• Ethics, data privacy, and responsible deployment of agentic AI
• Ethical implications of generative AI

• Demonstrate the underlying theory behind language modelling and embeddings, and highlight their role in LLMs.
• Describe the components of an agent harness and the function of each component in the perception-reasoning-action-observation loop.
• Compare tool use, memory strategies, and multi-agent orchestration patterns, and justify design choices for a given agentic system requirement.
• Examine ethics and policies on generative AI.
• Evaluate the potential of large language models (LLMs) in solving specific problems.
• Design and execute an evaluation suite that measures correctness, reliability, and regression across agent behaviours.
• Evaluate and mitigate potential risks to ensure the safety of AI systems in various workplace scenarios.