This course covers the engineering of modern AI systems built around foundation models. It begins with the practical use of pre-trained LLMs through APIs and open-weight models, including prompt and context engineering, structured outputs, and function calling. The first half of the semester focuses on knowledge augmentation: text embeddings, vector databases (FAISS, Qdrant, pgvector), dense and hybrid retrieval, rerankers, chunking strategies, and the design and evaluation of Retrieval-Augmented Generation (RAG) pipelines, including retrieval metrics, RAGAS, and LLM-as-judge methodologies.

The second half addresses agentic systems: tool use, planning, ReAct and plan-and-execute architectures, multi-agent orchestration, state and memory management. Throughout the semester, students study cross-cutting concerns including evaluation, observability and tracing, latency and cost optimization, parameter-efficient adaptation (LoRA, DPO) as an alternative to in-context approaches, and security topics such as prompt injection, data exfiltration, and guardrails. The course concludes with deployment considerations and a team project in which students build and evaluate a domain-specific agentic application over a real knowledge base. The course is taught in English.