The complete path.

Fourteen modules that build the complete mental model — what LLMs actually do, transformers without the math, RAG, agents, evals, security, and LLMOps. If you only read one thing, read this.

Build the mental model every other section depends on — what GenAI actually does, where the AI Engineer role sits, and the capability limits every practitioner must internalize before writing a single line of production code.

The retrieval foundation layer: how embeddings encode meaning, which similarity metrics are safe to use on which models, how ANN indexes trade recall for speed, and the 2026 vector database landscape with honest cost and scale guidance.

Where most LLM cost and latency lives. Engineers who understand prefill vs decode phases, PagedAttention, continuous batching, speculative decoding, and quantization formats can cut per-token cost by 2-5x and halve p95 latency — without changing a single model weight.