AI Foundry:
Foundations of GenAI Systems

About This Course

This course provides a rigorous foundation in generative AI systems, examining both their capabilities and inherent limitations. Students will explore the architectural principles and training methodologies behind large language models, investigate mechanisms of in-context learning and prompt sensitivity, analyze retrieval-augmented generation systems, and examine tool-calling and reasoning architectures.

The course emphasizes mechanistic understanding over superficial surveying, developing students' ability to decompose system behavior, characterize failure modes (including confabulations, retrieval issues, and compositional errors), and evaluate industry claims. Through methodological analysis and practical implementation, students will acquire the technical foundation necessary to contribute to the development of robust GenAI systems.

What You'll Learn

• 🏗️ LLM Architecture & Training Methodologies Architectural principles behind large language models and the training methodologies that shape their capabilities and limitations.
• 💬 In-Context Learning & Prompt Sensitivity Mechanisms of in-context learning, how models respond to prompt variations, and the implications for reliable system design.
• 🔍 Retrieval-Augmented Generation (RAG) Analysis of RAG system design: retrieval quality, grounding, and failure modes including retrieval errors and hallucination under poor context.
• 🤖 Tool-Calling & Reasoning Architectures How LLMs invoke external tools, chain reasoning steps, and where these architectures break down under compositional or multi-step demands.
• ⚠️ Failure Mode Analysis Systematic characterization of confabulations, retrieval failures, and compositional errors — developing the ability to decompose and predict system misbehavior.
• 🔬 Evaluating Industry Claims Methodological frameworks for critically assessing benchmarks, capability claims, and published results in the GenAI literature and industry.

Prerequisites

The following are desirable but not strictly required:

This course is open to all Purdue students.