Roadmap coverage map: every node → where it's taught
A node-by-node audit of the roadmap mind map against Course 02.
The "Where" column names the module that teaches each node (a "(notes)" tag means it is covered
in that module's notes rather than its lab). This course is the applied beginner path: depth on
the build skills, survey-level on the rest; every node below is at least present.
Introduction
| Node |
Where |
| What is an AI Engineer · AI Engineer vs ML Engineer · AI vs AGI |
M0 |
| Impact on product development · Using AI to improve UX |
M0 (notes) |
How LLMs work
| Node |
Where |
| LLMs · How LLMs work · Next-token prediction · Inference · Tokenization |
M0 (+ tokens M6) |
| Understanding model capabilities |
M0 (model landscape) |
| Neural networks · Transformers |
M0 (go-deeper: "how LLMs are built") |
Models & selection
| Node |
Where |
| Pre-trained · Closed vs open · Choosing / model selection · Model families · Smaller models |
M0 (+ M13) |
| SKU / variants · Understanding capabilities |
M0 / M6 |
Using model APIs
| Node |
Where |
| Claude Messages API · Input/Output format · Structured outputs · System prompts |
M4, M5, M6 |
Token counting (count_tokens) |
M6 (go-deeper) |
| OpenAI API · Google Gemini API · Hugging Face Inference SDK |
M6 (cross-provider note) / M13 |
Prompt engineering
| Node |
Where |
| Prompt engineering · System prompts · Few-shot · CoT · Constraining I/O |
M5 (+ M10) |
| Context engineering · Prompt optimization · Prompt compression |
M5 (notes) |
AI safety & ethics
| Node |
Where |
| AI safety & ethics · Bias & fairness · Content moderation |
M14 (+ M10) |
| Conducting adversarial testing · Safety evaluation |
M10 |
| Data classification · Anomaly/abuse detection · End-user IDs · Know-your-customer |
M10 (operational controls) |
Open-source & local models
| Node |
Where |
| Open-source models · Ollama · LM Studio · Hugging Face (hub/models/tasks/SDK) · local/remote server · Quantization |
M13 |
Embeddings
| Node |
Where |
| Embeddings · Vector embeddings · Indexing · Semantic search |
M7 |
| Embedding models / providers (OpenAI, Cohere, Gemini, Voyage, Jina) |
M7 (go-deeper) |
Vector databases
| Node |
Where |
| Vector databases · Chroma · FAISS |
M7 |
| LanceDB · Pinecone · Qdrant · Weaviate (+ pgvector) |
M7 (go-deeper) |
RAG
| Node |
Where |
| RAG · Chunking · Indexing · Retrieval · Ranking · Re-ranking · Generation · Data layer · Semantic search · RAG chatbots · RAG over multimodal |
M7, M8 (multimodal RAG: M12) |
| Agentic RAG · retrieval as a tool · multi-hop · query refinement · citations · research agents |
M24 (plain-vs-agentic RAG, multi-hop, verified offline) |
AI agents
| Node |
Where |
| What are agents · Use cases · Function calling · Tools · Manual loop · ReAct · External memory |
M9 (+ M21 for memory) |
| Agent memory & state · short-term / long-term memory · token budget · checkpoint & resume |
M21 (built & verified offline: memory store + memory agent) |
| Agent reliability & ops · retries/backoff · timeouts · fallback · step caps · approval gates |
M22 (built & verified offline via fault injection) |
| Agent security · prompt injection (indirect) · tool poisoning · excessive agency · exfiltration · least privilege · defense in depth |
M23 (vulnerable-vs-hardened demo, verified offline; synthetic/defensive) |
| Context compaction · Context isolation |
M9 (notes) + M21 (token budget, summarize-oldest) |
| Frameworks: LangChain · LlamaIndex · Haystack · dev tools |
M9 (LangGraph deep + survey) |
| Coding agents: Claude Code · Cursor · OpenAI Codex · Google ADK · Claude Agent SDK |
M9 (survey addition) |
| Multi-agent systems · orchestrator/coordinator · sub-agents · hand-offs · connectors |
M18 (deployed SOC orchestrator, built & verified) |
| Orchestration shapes: pipeline · router · parallel fan-out · hierarchical |
M18 (pipeline built; others explained, built with frameworks in M19) |
| Agent frameworks breadth: LangChain · LangGraph · CrewAI · AutoGen · OpenAI Agents SDK · Claude Agent SDK · smolagents · LlamaIndex · n8n |
M19 (the same agent built across all 8 + no-code n8n, built & verified) |
| Complete agent (capstone): RAG + memory + observability + reliability + security + eval gate + API |
M27 (Part D capstone integrating M18-M26, verified offline + TestClient) |
| Agent UX · streaming · perceived latency · progress/citations/cost · cancellation · Server-Sent Events |
M28 (event-stream agent + SSE endpoint, verified offline + TestClient) |
MCP
| Node |
Where |
| What is MCP · MCP client · MCP server |
M9 (concept) + M16 |
| Building an MCP client / server |
M16 (full build, verified end-to-end) |
Multimodal AI
| Node |
Where |
| Multimodal · Image understanding · Vision-language models |
M12 |
| Image generation · DALL·E · Audio · Speech recognition · Whisper · Video · Multimodal RAG · LangChain/LlamaIndex multimodal |
M12 (survey) |
Classic NLP tasks
| Node |
Where |
| NLP tasks · Sentiment analysis · Summarization · Data classification |
M5 (these are prompting tasks, notes + examples) |
| Anomaly detection |
M10/M14 (abuse/anomaly) |
| Web search integration |
M9 (agent tool) |
Evaluation, optimization & monitoring
| Node |
Where |
| Model evaluation · Safety evaluation |
M8 (RAG), M14 (fairness), M10 (safety) |
| Agent evaluation · eval harness · scorers · golden set · LLM-as-judge · regression tests |
M20 (built & verified: traced agent + scorecard) |
| Evaluation-driven development · CI eval gate · GitHub Actions · quality tracking |
M26 (gate + exit codes + sample workflow, verified offline) |
| Data flywheel · feedback capture · curation into evals + fine-tuning data · PII redaction |
M30 (feedback log + curation, verified offline; ties M26 + M15 + M14) |
| Observability · tracing · spans · OpenTelemetry · cost/latency/error monitoring |
M20 (built a tracer; survey of LangSmith/Langfuse/Phoenix) |
| Monitoring · Monitoring LLM apps · Inference cost/latency |
M11 (+ M6, M20) |
| Deployment & serving · containers · env config/secrets · health/readiness probes · stateless scaling |
M11 (+ M29 production serving: probes, lifecycle, statelessness, Dockerfile) |
| Cost & performance optimization · prompt caching · model routing · token trimming · batching |
M25 (offline cost/latency estimator, verified) |
| Model optimization · Quantization |
M13 |
Fine-tuning & training
| Node |
Where |
| Fine-tuning (SFT) · Training custom models · RLHF · LoRA/PEFT · dataset prep |
M15 (full module + hands-on dataset build) |
| Transformers · Neural networks · how training works |
M0 (gist) + M15 (pipeline) + M17 (build one from scratch) |
| Train a transformer/LLM from scratch |
M17 (tiny LM in numpy, verified; miniature GPT in PyTorch) |
Data & integrations
| Node |
Where |
| SQL databases · Web search integration · Data layer |
M9 (agent tools), M8 (data layer) |
Now fully built (per course-owner direction): fine-tuning & training is a complete hands-on
module (M15), dataset building, the fine-tune workflow (hosted + local LoRA), and the
when-to-fine-tune decision. Every roadmap branch is taught or built.
Every roadmap branch is now taught or built, nothing is left as "out of scope." The
fine-tuning/training branch is a full hands-on module (M15), building MCP servers/clients is a
full, end-to-end-verified module (M16), and training a transformer/LLM from scratch is now a
hands-on deep-dive too (M17: a tiny model trained from scratch in numpy, verified for real, plus a
miniature PyTorch transformer). M17 is flagged optional / "researcher" lab, understanding, not the
engineer's day job, but it's there, built, for anyone who wants to see all the way down.