Sources
The roadmap’s courses, books, articles and repos are listed here with their purpose, perspective and date.
- Recommendation review (2026-11-09): course lectures, textbook chapter numbers and assignment topics were checked against original syllabi, tables of contents and assignment descriptions. This does not mean every video was watched, every external assignment completed, or examples run on every hardware setup. External resources are optional; choose the sections relevant to your goal and follow their own prerequisites.
- Maintenance: this field moves fast, so the list is checked once a quarter. If a link is broken or the content is out of date, please report it or open a PR
- Chapters: the Chapters column lists which chapters each resource goes with; for how to read the codes before the chapter titles, see How to read the codes
- Language: entries marked (in Chinese) are in Chinese; everything else is in English
University courses
Section titled “University courses”Books and long-form material
Section titled “Books and long-form material”| Source | What it’s for | Chapters |
|---|---|---|
| Chip Huyen, AI Engineering (O’Reilly 2025), companion repo | The backbone of application engineering: evals, RAG, fine-tuning, data, inference, architecture, user feedback. A paid book; the repo has summaries only | S5 Inference and Serving, S7 Prompts, Structured Output and Workflows to S9 Retrieval and RAG, S17 Product and AI UX, S18 Deployment and Architecture |
| rasbt/LLMs-from-scratch | Implements GPT from scratch; bonus material covers KV cache, GQA, MLA, MoE, DPO | S2 Transformers and Decoding to S5 Inference and Serving |
| karpathy/nn-zero-to-hero | Video series; the best intuition for backprop and training. Not updated since 2024 | S1 Neural Network Basics to S2 Transformers and Decoding |
| mlabonne/llm-course | A roadmap-style map of links; unusually complete on data engineering, quantization and security. Content not updated since 2025-12 | S3 Pretraining and Scaling to S5 Inference and Serving, S9 Retrieval and RAG |
| Hugging Face Ultra-Scale Playbook | The standard reading on parallel training | S3 Pretraining and Scaling |
| Hugging Face LLM Course, MCP Course, Agents Course | Chapters 11–12 cover SFT and GRPO; the Agents Course leans on frameworks | S4 Post-training and Reasoning, S10 Tools and MCP, S12 The Agent Loop and Harness |
| microsoft/ai-agents-for-beginners | 18 lessons; L03 is one of the few that covers agentic UX. Heavily tied to Azure | S16 Security and Governance to S18 Deployment and Architecture |
| anthropics/courses | Prompting tutorials and a prompt evaluations lab. Archived, but the content is still usable | S7 Prompts, Structured Output and Workflows, S8 Evals Methodology |
| humanlayer/12-factor-agents | “Own your prompts, context and control flow.” Partly promotes HumanLayer’s product | S7 Prompts, Structured Output and Workflows, S11 Context Engineering, S12 The Agent Loop and Harness |
| SWE-agent/mini-swe-agent | A complete agent in about 100 lines | S0 AI Systems Overview, S12 The Agent Loop and Harness |
| WenyuChiou/awesome-agentic-ai-zh (in Chinese) | An agent learning map in Traditional Chinese; strong 8-component harness checklist and evals. Leans toward the Claude Code ecosystem; shallow on how models work | S12 The Agent Loop and Harness, S14 Agent Evals, Data and Optimization |
| karpathy/makemore | Character-level language model; the S1 Neural Network Basics exercise uses its names.txt data. Not updated since 2024 |
S1 Neural Network Basics |
| karpathy/nanoGPT | The leanest GPT training repo; S2 Transformers and Decoding borrows its config while you write the model yourself | S2 Transformers and Decoding |
| karpathy/nanochat | End-to-end training of a small ChatGPT; the full version needs 8×H100, with scaled-down CPU / MPS versions | S3 Pretraining and Scaling, model project |
| GeeeekExplorer/nano-vllm | A minimal vLLM in about 1,200 lines, to compare against your own inference server | S5 Inference and Serving, model project |
SWE-agent/mini-swe-agent: agents/default.py |
Source of the agent’s main loop; S0 AI Systems Overview walks through it method by method | S0 AI Systems Overview |
| Microsoft ai-agents-for-beginners L03: AI Agentic Design Principles | Three principles: transparency, control, consistency. Examples are tied to Azure; the principles are general | S17 Product and AI UX |
| Microsoft: Guidelines for Human-AI Interaction (2019) | 18 human-AI interaction guidelines, usable as a UI audit checklist | S17 Product and AI UX |
| Nielsen Norman Group: Response Times: The 3 Important Limits (Nielsen 1993, drawing on Miller 1968 and Card et al. 1991) | The three response-time thresholds: 0.1 s / 1 s / 10 s | S17 Product and AI UX |
| roadmap.sh AI Engineer, AI Agents | A coverage checklist; a thorough summary of context failure modes. Shallow, mostly lists of tools | All |
Industry articles and guides
Section titled “Industry articles and guides”Anthropic Engineering
Section titled “Anthropic Engineering”OpenAI, Google / Kaggle, Meta
Section titled “OpenAI, Google / Kaggle, Meta”Individual authors
Section titled “Individual authors”Security standards
Section titled “Security standards”Official docs and tools
Section titled “Official docs and tools”API docs, SDKs, libraries, model cards, datasets, and local inference and evaluation tools change quickly. Check the current documentation before using them.
Claude API
Section titled “Claude API”| Source | What it’s for | Chapters |
|---|---|---|
| Claude API: Messages | Message structure | S6 LLM APIs and Application Basics |
| Claude API: Streaming messages | Order of SSE events | S6 LLM APIs and Application Basics |
| Claude API: Errors | Which errors to retry | S6 LLM APIs and Application Basics |
| Claude API: Rate limits | Rate limits | S6 LLM APIs and Application Basics |
| Claude API: Pricing | Pricing | S6 LLM APIs and Application Basics |
| Claude API: Batch processing | Batch API | S6 LLM APIs and Application Basics |
| Claude API: Thinking | Reasoning model parameters and passing thinking back across turns | S6 LLM APIs and Application Basics |
| Claude API: Prompt engineering overview | Official prompting advice | S7 Prompts, Structured Output and Workflows |
| Claude API: Structured outputs | Structured output | S7 Prompts, Structured Output and Workflows |
| Claude API: Tool search tool | The official implementation of deferred loading | S10 Tools and MCP |
| Claude API: Prompt caching | Pricing, minimum length, lifetime, usage fields | S11 Context Engineering |
OpenAI API
Section titled “OpenAI API”| Source | What it’s for | Chapters |
|---|---|---|
| OpenAI API: Streaming API responses | Streaming events | S6 LLM APIs and Application Basics |
| OpenAI API: Error codes | Error codes | S6 LLM APIs and Application Basics |
| OpenAI API: Reasoning models | Reasoning model parameters, pricing | S6 LLM APIs and Application Basics |
| OpenAI API: Structured model outputs | Structured output | S7 Prompts, Structured Output and Workflows |
| OpenAI API: Prompt caching | Prompt caching | S11 Context Engineering |
| Source | What it’s for | Chapters |
|---|---|---|
| MCP specification 2026-07-28: Overview | Host / client / server, feature list, security principles | S10 Tools and MCP |
| MCP specification: Architecture | Architecture | S10 Tools and MCP |
| MCP specification: Authorization | Authorization | S10 Tools and MCP |
| MCP specification: Transports | Transport layer | S10 Tools and MCP |
| MCP specification: Changelog | Differences between versions | S10 Tools and MCP |
| MCP: Security Best Practices | Security practice; redirects to the 2026-07-28 version | S10 Tools and MCP |
| modelcontextprotocol/typescript-sdk | Official TypeScript SDK | S10 Tools and MCP |
| modelcontextprotocol/python-sdk | Official Python SDK | S10 Tools and MCP |
| FastMCP | Python MCP framework | S10 Tools and MCP |
| MCP Inspector | Testing tool for MCP servers | S10 Tools and MCP |
| modelcontextprotocol/ext-apps (MCP Apps) | Official extension for MCP servers to provide UI; stable spec 2026-01-26 | S17 Product and AI UX |
| MCP-UI | The community project and SDK that preceded MCP Apps | S17 Product and AI UX |
SDKs, frameworks and protocols
Section titled “SDKs, frameworks and protocols”Local models and inference
Section titled “Local models and inference”| Source | What it’s for | Chapters |
|---|---|---|
| vLLM | Source code of a production-grade inference engine | S5 Inference and Serving |
| Ollama API: Usage | Timing fields such as prompt_eval_count and eval_duration. Replaces docs/api.md on GitHub |
S5 Inference and Serving |
| Ollama: OpenAI compatibility | OpenAI-compatible interface for local models | S6 LLM APIs and Application Basics |
| llama.cpp server | Local HTTP server | S6 LLM APIs and Application Basics |
| mlx-lm | Run LLMs on Apple Silicon | S6 LLM APIs and Application Basics |
| transformers.js | Run models in the browser | S6 LLM APIs and Application Basics, S9 Retrieval and RAG |
| WebLLM | In-browser LLM inference engine | S6 LLM APIs and Application Basics |
| Chrome built-in AI | APIs for the browser’s built-in models | S6 LLM APIs and Application Basics |
Training tools
Section titled “Training tools”| Source | What it’s for | Chapters |
|---|---|---|
| Weights & Biases | Experiment tracking; has a free personal plan. The docs are now part of CoreWeave | S1 Neural Network Basics, model project |
| Hugging Face TRL | SFT, DPO and GRPO trainers | S4 Post-training and Reasoning, model project |
| Hugging Face PEFT | LoRA and other parameter-efficient fine-tuning | S4 Post-training and Reasoning, model project |
| Hugging Face: Chat templates | Read the section on tool use | S7 Prompts, Structured Output and Workflows |
Models and datasets
Section titled “Models and datasets”| Source | What it’s for | Chapters |
|---|---|---|
| Qwen/Qwen3-0.6B-Base | Starting point for the post-training exercises | S4 Post-training and Reasoning |
| Ollama: qwen3 | Run Qwen3 locally | S4 Post-training and Reasoning |
| openai/gsm8k | Math word problem dataset; likely contaminated, so use it only for relative comparisons | S4 Post-training and Reasoning |
| Qwen/Qwen3-0.6B | Small instruct model; in S5 Inference and Serving you write your own generation loop for it | S5 Inference and Serving |
| Qwen/Qwen3-Embedding-0.6B | An embedding model that runs on your own machine | S9 Retrieval and RAG |
| BAAI/bge-m3 | Multilingual embedding model | S9 Retrieval and RAG |
| BAAI/bge-reranker-v2-m3 | Multilingual reranker | S9 Retrieval and RAG |
| MTEB Leaderboard | Reference for choosing an embedding model | S9 Retrieval and RAG |
| tiny Shakespeare | About 1.1 MB of plain English text; training data for the S2 Transformers and Decoding core exercise and M2 Tiny GPT | S2 Transformers and Decoding |
| HuggingFaceFW/fineweb | About 15T tokens of cleaned English web data; the large-scale deduplication extension in S3 Pretraining and Scaling uses a sample of it | S3 Pretraining and Scaling |
| Salesforce/wikitext | Wikipedia articles; the document source for deduplication in the S3 Pretraining and Scaling core exercise | S3 Pretraining and Scaling |
| Qwen3-0.6B config.json | What the inference calculator is checked against | S5 Inference and Serving |
| Spider | Text-to-SQL dataset (200 databases); the ○ Text2SQL variant in S9 Retrieval and RAG can use a subset | S9 Retrieval and RAG |
| Laws & Regulations Database of the Republic of China (Taiwan) (in Chinese) | Full text of Taiwanese laws in Traditional Chinese; the document set for the S9 Retrieval and RAG core exercise | S9 Retrieval and RAG |
| LongMemEval | Data and code for long-term memory evals; paper: Wu et al. 2024 | S13 Memory and Personalization |
| MultiHop-RAG | QA set whose questions need several news articles combined; paper: Tang & Yang 2024; ODC-BY license | S15 Multi-agent Systems |
| abisee/cnn_dailymail | English news articles and summaries; the English alternative dataset for the S8 Evals Methodology core exercise (version 3.0.0) | S8 Evals Methodology |
| goodbooks-10k | Metadata and ratings for ten thousand books; the data for the S10 Tools and MCP core exercise | S10 Tools and MCP |
RAG and retrieval
Section titled “RAG and retrieval”| Source | What it’s for | Chapters |
|---|---|---|
| Sentence Transformers | Embeddings and cross-encoders | S9 Retrieval and RAG |
| Sentence Transformers: Training Overview | Official tutorial on fine-tuning embeddings | S9 Retrieval and RAG |
| SQLite FTS5 | BM25 full-text search; for Chinese text, use trigram or segment words first | S9 Retrieval and RAG |
| jieba | Chinese word segmentation for Python | S9 Retrieval and RAG |
| sqlite-vec | Vector search extension for SQLite | S9 Retrieval and RAG |
| LanceDB | Embedded vector database | S9 Retrieval and RAG |
| unpdf | PDF text extraction for JavaScript | S9 Retrieval and RAG |
| Docling | Parsing documents with complex layouts | S9 Retrieval and RAG |
| Ragas | Definitions of RAG evaluation metrics | S8 Evals Methodology, S9 Retrieval and RAG |
Evaluation, testing and observability
Section titled “Evaluation, testing and observability”| Source | What it’s for | Chapters |
|---|---|---|
| OpenTelemetry GenAI semantic conventions | Standard trace fields for LLM calls; still in Development status | S6 LLM APIs and Application Basics |
| OpenLLMetry | LLM observability built on OpenTelemetry | S6 LLM APIs and Application Basics |
| Langfuse | Open-source LLM tracing and eval platform | S6 LLM APIs and Application Basics |
| Phoenix | Open-source LLM observability and evaluation | S6 LLM APIs and Application Basics |
| Jaeger | A plain tracing backend | S6 LLM APIs and Application Basics |
| MSW (Mock Service Worker) | API mocking for TypeScript | S6 LLM APIs and Application Basics |
| httpx | Python HTTP client | S6 LLM APIs and Application Basics |
| respx | Mocking for httpx | S6 LLM APIs and Application Basics |
| pytest-recording | Record / replay built on VCR.py | S6 LLM APIs and Application Basics |
| promptfoo | Eval runner | S8 Evals Methodology |
scipy.stats.bootstrap |
Bootstrap confidence intervals | S8 Evals Methodology |
| AgentDojo | An evaluation environment for prompt injection attacks and defenses; paper: Debenedetti et al. 2024. The S16 Security and Governance core exercise borrows its injection templates | S16 Security and Governance |
Web and API basics
Section titled “Web and API basics”| Source | What it’s for | Chapters |
|---|---|---|
| MDN: Using server-sent events | The SSE format itself | S6 LLM APIs and Application Basics |
| AWS Builder Center: Timeouts, retries, and backoff with jitter | Engineering practice for retries and backoff | S6 LLM APIs and Application Basics |
| Stripe: Designing robust and predictable APIs with idempotency | Designing idempotency keys | S6 LLM APIs and Application Basics |
| IETF draft: The Idempotency-Key HTTP Header Field | draft-08, not yet an RFC | S6 LLM APIs and Application Basics |
Papers
Section titled “Papers”Every paper cited in the chapters’ Further reading is listed here.
Chinese-language resources
Section titled “Chinese-language resources”Traditional Chinese
Section titled “Traditional Chinese”The iThome series are 30-day posts by practitioners, with varying quality. Those listed above won awards or closely match the topic. Use them alongside a primary resource.
Simplified Chinese
Section titled “Simplified Chinese”| Source | What it’s for | Chapters |
|---|---|---|
| RUC: Large Language Models (in Chinese) | A university textbook, free PDF; the agent and application parts are dated | S0 AI Systems Overview, S3 Pretraining and Scaling, S4 Post-training and Reasoning |
| datawhalechina/diy-llm (in Chinese) | A Chinese version of CS336; the best Chinese companion to Part I Models (S1–S5) | S2 Transformers and Decoding to S5 Inference and Serving, S8 Evals Methodology |
| datawhalechina/happy-llm (in Chinese) | Builds an LLM from scratch; chapter 8 goes deep on agentic RL | S2 Transformers and Decoding, S4 Post-training and Reasoning, S14 Agent Evals, Data and Optimization |
| ZJU-LLMs/Foundations-of-LLMs (in Chinese) | Zhejiang University’s Foundations of LLMs; has a chapter on model editing | S4 Post-training and Reasoning, S7 Prompts, Structured Output and Workflows, S9 Retrieval and RAG |
| Lordog/dive-into-llms (in Chinese) | Shanghai Jiao Tong University’s Dive into LLMs; many security labs | S4 Post-training and Reasoning, S16 Security and Governance |
| datawhalechina/tiny-universe (in Chinese) | Minimal implementations: TinyTransformer / TinyEval / TinyRAG / TinyAgent | S2 Transformers and Decoding, S8 Evals Methodology, S9 Retrieval and RAG, S12 The Agent Loop and Harness |
| datawhalechina/zero-to-sglang (in Chinese) | A course made with the SGLang team: write your own inference server. Part three of the course is not finished yet | S5 Inference and Serving, S18 Deployment and Architecture |
| datawhalechina/all-in-rag (in Chinese) | The best Chinese material for S9 Retrieval and RAG; leans on Milvus | S9 Retrieval and RAG |
| datawhalechina/hello-agents (in Chinese) | Complete material for building agents from scratch; deep on context engineering, memory and agentic RL. Leans toward the Chinese tech stack | S11 Context Engineering to S15 Multi-agent Systems |
| datawhalechina/agentic-rl-lab (in Chinese) | Runnable reproductions of GRPO variants, Search-R1 and ReTool | S14 Agent Evals, Data and Optimization |
| datawhalechina/base-llm (in Chinese) | Covers security, threat modeling and deployment; one chapter is a tutorial for a vendor’s API | S16 Security and Governance, S18 Deployment and Architecture |
| datawhalechina/deepagents-in-action (in Chinese) | A complete topic map of harnesses; tied to LangChain | S12 The Agent Loop and Harness, S13 Memory and Personalization, S15 Multi-agent Systems |
| shareAI-lab/learn-claude-code (English, with a Chinese translation) | 17 lessons that hand-write a Claude Code-style harness from scratch; the main material for S12 The Agent Loop and Harness. Follow s01–s17 in the repo root; the old docs/ folder is numbered differently |
S10 Tools and MCP to S15 Multi-agent Systems, S19 Real Agent Systems and the Capstone |
| d2l-zh: Dive into Deep Learning (in Chinese) | Deep learning basics; not updated since 2024 | S1 Neural Network Basics to S2 Transformers and Decoding |
| youngyangyang04/llm-master (in Chinese) | A full-stack LLM learning path; its engineering acceptance criteria are a useful reference. Includes links to a paid bootcamp and top-up services; heavily job-hunting oriented | S18 Deployment and Architecture |
| datawhalechina/self-llm (in Chinese) | Fine-tuning and deployment recipes for many models | S4 Post-training and Reasoning, S18 Deployment and Architecture |
| datawhalechina/llm-cookbook (in Chinese) | Chinese version of Andrew Ng’s short courses; “Building Systems with the ChatGPT API” works as an API warm-up. Not updated since 2025-06 | S6 LLM APIs and Application Basics |
Lecture content is licensed under CC BY-NC-SA 4.0: you may share and adapt it with attribution, not for commercial use, and adaptations must use the same license. Quoted third-party content and trademarks belong to their owners.