The gap between a research result and a production system is enormous. This course bridges it by covering the four pillars of applied AI systems: retrieval and knowledge grounding, autonomous agents, evaluation and safety, and production engineering.

The retrieval section goes from vector search algorithms (HNSW, IVF, product quantization) through basic RAG to advanced techniques like query decomposition, self-RAG, and graph RAG. You will build a complete RAG pipeline from scratch and evaluate it with RAGAS.

The agents section covers prompting, tool use, and agent architectures (ReAct, Reflexion, multi-agent debate). You will build a ReAct agent from scratch and learn context engineering — the art of managing what goes into the context window.

The evaluation and safety section tackles the hardest unsolved problems: how to evaluate LLMs reliably, how to interpret what is happening inside them (mechanistic interpretability), and how to prevent them from doing harmful things. These are the problems that define the research frontier.

The production section covers inference optimization (quantization, speculative decoding, continuous batching), distributed training (DDP, ZeRO, FSDP), and the meta-skill of reading and reproducing research papers.