Goal
Everything an engineer needs to build, evaluate, secure, and operate production LLM systems — from how a model decodes a token to agentic systems, RAG, fine-tuning, evals, and LLMOps. Concept-first and tool-aware, with animated diagrams for the ideas that are easier to see than to read. Pairs with LangChain Senior Prep / LangGraph Senior Prep (the frameworks) and FastAPI Senior Prep (serving it).
How to use this track
You don’t build LLM features by knowing one big thing — you combine many small, sharp ideas correctly. Go top-to-bottom for a full pass, or jump to the module you need. Every module follows: concept → how it works → pros/cons → caveats → risk mitigation → in practice → sources.
This field churns — separate the durable from the volatile
The concepts here are stable (sampling reshapes a distribution, retrieval is nearest-neighbour, evals gate releases). The numbers and APIs are not — model IDs, prices, cache thresholds, framework versions, and even which sampling params exist change every few months. Each module flags its high-churn facts; re-verify those against the primary source before you rely on them. Today’s date context: mid-2026.
Three concerns that run through EVERY module (the spine)
- Security — prompt injection (esp. indirect, via retrieved/tool content) is not solved by RAG or fine-tuning. See 14 - Safety, Security & Risk.
- Cost & latency — input ≠ output tokens; agents cost ~4×, multi-agent ~15× a chat call. The real metric is cost-per-successful-task. See 15 - Cost, Latency & Performance.
- Evaluation — the component is non-deterministic, so every change needs offline evals. See 13 - Evaluation & Evals.
Curriculum
| # | Module | Core question | Animated? | Status |
|---|---|---|---|---|
| 01 | 01 - How LLMs Work | tokens, context, attention, autoregressive decoding, KV cache | 🎞️ | ready |
| 02 | 02 - Decoding, Sampling & Reasoning Controls | temperature, top-p/k, penalties, logprobs, reasoning_effort | 🎞️ | ready |
| 03 | 03 - Prompt & Context Engineering | roles, few-shot, CoT, prompt caching, context rot | ready | |
| 04 | 04 - Structured Output & Tool Calling | JSON schema, constrained decoding, function calling | ready | |
| 05 | 05 - Model Context Protocol (MCP) | the tool/integration standard for agents | ready | |
| 06 | 06 - Embeddings & Vector Search | embeddings, ANN, hybrid search, reranking | 🎞️ | ready |
| 07 | 07 - RAG | chunk → retrieve → rerank → augment → generate | 🎞️ | ready |
| 08 | 08 - Agents & Agentic Systems | ReAct, planning, reflection, workflow-vs-agent | 🎞️ | ready |
| 09 | 09 - Workflow & Orchestration Patterns | chaining, routing, parallel, orchestrator, evaluator | ready | |
| 10 | 10 - Multi-Agent Systems | supervisor, swarm, handoffs, when it helps vs hurts | 🎞️ | ready |
| 11 | 11 - Memory for LLM Apps | short/long-term, agentic memory, forgetting | ready | |
| 12 | 12 - Fine-tuning & Model Adaptation | SFT, LoRA/QLoRA, DPO/GRPO, distillation | 🎞️ | ready |
| 13 | 13 - Evaluation & Evals | LLM-as-judge, RAGAS, eval-driven dev, agent eval | 🎞️ | ready |
| 14 | 14 - Safety, Security & Risk | injection, OWASP LLM Top 10, guardrails, red-teaming | 🎞️ | ready |
| 15 | 15 - Cost, Latency & Performance | caching, routing/cascades, quantization, batching | ready | |
| 16 | 16 - Observability & LLMOps Lifecycle | tracing, prompt versioning, drift, the lifecycle loop | 🎞️ | ready |
| 17 | 17 - Deployment & Serving | vLLM/SGLang, GPU sizing, provider vs self-host | ready | |
| 18 | 18 - Multimodal, Voice & Computer-Use | vision/audio RAG, speech agents, computer-use | ready | |
| 19 | 19 - Agent Protocols & Interop | MCP · A2A · AG-UI · A2UI · AP2 — the interop stack | 🎞️ | ready |
Suggested order
Foundations (01–05) → Retrieval (06–07) → Agents (08–11) → Adapting & proving (12–13) → Operating safely & cheaply (14–17) → Frontier modalities (18). If you only do six: 02, 03, 07, 08, 13, 14 — the daily-driver core of building LLM apps that actually work and don’t get owned.
How this fits the rest of the vault
PYTHON MASTERY + PYTHON ADVANCED the language + runtime
+
FASTAPI SYSTEMS MASTERY serving APIs at scale
+
LANGCHAIN / LANGGRAPH PREP the orchestration frameworks
▼
GEN AI ENGINEER (this) the LLM-specific craft: decoding, RAG,
agents, fine-tuning, evals, safety, LLMOpsFramework how-tos live in LangChain Senior Prep and LangGraph Senior Prep; this track is framework-agnostic concepts you’d apply with any of them (or the raw provider SDK).
Anchor sources (re-verify, they move)
- Anthropic: Building Effective Agents · Effective Context Engineering · Contextual Retrieval · Multi-Agent Research System
- OpenAI: Reasoning guide · Structured Outputs · Model Spec
- Safety: OWASP GenAI LLM Top 10 (2025) · NIST AI RMF
- Eval: RAGAS · key papers: ReAct, RAG, LoRA, DPO
- Standard: Model Context Protocol
On the diagrams
Animated GIFs are generated by
assets/gen_genai_gifs.py(Pillow). Re-run it to tweak. Obsidian animates GIFs in reading view (animated SVG renders frozen, so GIF is used).