Tokens, transformers-at-a-glance, and why models hallucinate — the conceptual base for everything else.

Providers, open vs closed models, and choosing by capability, cost, and latency.

Getting reliable behavior from models: structure, examples, and iteration.

Messages, roles, parameters, and handling API failures properly.

Getting JSON you can actually parse: schemas, validation, and repair strategies.

Semantic vectors: what they capture, and similarity math.

Using vector stores in applications: indexes, filters, and choosing one. (Capacity math and scaling live in System Design.)

The full loop: ingest → chunk → embed → retrieve → generate, with citations.

Query rewriting, multi-hop retrieval, and agentic RAG.

The agent loop: reason, act, observe — and when an agent is overkill.

Planning, reflection, and multi-agent designs that actually help.

Making AI features fast and affordable at scale.

Input/output filtering, prompt injection, and responsible defaults.

Curating the window: retrieved docs, tool schemas, history, and constraints — and the failure modes of too much.

Role, constraints, and format as a designed artifact — versioned and injection-resistant.

LangGraph vs AI SDK vs Agents SDKs — and what a harness actually owns.

Designing tools, resources, and prompts that agents can actually use well.

The unglamorous 80%: PDFs, OCR, tables, and metadata enrichment.

Knowledge-graph retrieval: when relationships beat chunks.

The decision tree: prompting vs RAG vs fine-tuning, with honest cost math.

Parameter-efficient tuning: adapters, data prep, and what actually matters.

vLLM and friends: continuous batching and KV-cache-aware capacity planning.

Smaller, faster models — and testing that quality survived.

Speech-to-speech loops: latency budgets, barge-in, and telephony.

Document VQA, UI understanding, and image inputs done reliably.

The security checklist applied to a real LLM application.

Input/output classifiers, policy models, and jailbreak monitoring.