翻訳待ち：A Markdown curriculum for engineers moving into AI engineering

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

Notifications You must be signed in to change notification settings Fork 0 Star 1 BranchesTags Open more actions menu Folders and files NameName Last commit message Last commit date Latest commit History 16 Commits 16 Commits 01-llm-fundamentals 01-llm-fundamentals 02-prompting-and-context-engineering 02-prompting-and-context-engineering 03-retrieval-and-rag 03-retrieval-and-rag 04-agents-and-tool-use 04-agents-and-tool-use 05-evaluation-and-observability 05-evaluation-and-observability 06-deployment-and-ai-infra 06-deployment-and-ai-infra 07-emerging-topics 07-emerging-topics assets assets .gitignore .gitignore LICENSE LICENSE README.md README.md capstone.md capstone.md hot-topics.md hot-topics.md soft-skills.md soft-skills.md Repository files navigation A structured, opinionated curriculum for software engineers moving into AI/ML engineering — the discipline that separates someone who can call an LLM API from someone who can design, evaluate, ship, and operate AI systems that hold up in production. It's built as plain markdown: seven modules from foundations to frontier, each with canonical resources, the tools actually used in industry, a hands-on project, and the pitfalls that bite people first. I'm writing it as I learn (publicly, mistakes included), and it's open for anyone making the same transition — self-taught devs, backend/full-stack engineers adding AI to their scope, and new grads who want depth beyond "I used the OpenAI SDK once." Last reviewed: August 2026 · See hot-topics.md for what's moving fast vs. what's settled. Table of Contents # Module What you'll be able to do 01 LLM Fundamentals Explain what the model is actually doing, and pick one on evidence 02 Prompting & Context Engineering Get reliable, structured output and manage the context window as a budget 03 Retrieval & RAG Build retrieval that measurably finds the right thing 04 Agents & Tool Use Design agent loops that terminate, recover, and stay in budget 05 Evaluation & Observability Replace vibes with numbers, offline and in production 06 Deployment & AI Infra Serve models at a known cost, latency, and failure profile 07 Emerging Topics Evaluate new capabilities without chasing demos Cross-cutting files hot-topics.md — what changed in the last 6–12 months vs. stable fundamentals soft-skills.md — reviewing AI-generated code, cost/latency tradeoffs, security awareness, communicating limitations capstone.md — the final project that combines every module Calling the model is the small box in the middle. This repo is about everything around it. Progress Tracker Fork the repo and tick these off as you go. A box is "done" when you can explain the idea to another engineer and you've shipped code that uses it. 01 · LLM Fundamentals Transformers and the attention mechanism Tokenization and the context window The training lifecycle: pretraining → SFT → preference tuning Inference and decoding: sampling params, KV cache, streaming Model selection, scaling laws, and reasoning models Project: model bake-off harness 02 · Prompting & Context Engineering Prompt anatomy and instruction following Few-shot examples and reasoning elicitation Structured output and schema enforcement Context engineering: assembly, compaction, and caching Prompt versioning, testing, and optimization Project: structured extraction pipeline 03 · Retrieval & RAG Embeddings and vector search Chunking, parsing, and index design Hybrid search and reranking Query understanding: rewriting, routing, filtering RAG evaluation and failure modes Project: cited answers over a real corpus 04 · Agents & Tool Use Tool calling and the agent loop Agent architectures: workflows vs. agents Memory and state management Environments and protocols (MCP, sandboxes, computer use) Reliability: budgets, retries, human-in-the-loop Project: bounded research agent with durable state 05 · Evaluation & Observability Eval fundamentals and error analysis LLM-as-judge and judge calibration Tracing, metrics, and OTel GenAI conventions Production feedback loops and CI regression gates Agent and RAG-specific evaluation Project: eval harness wired into CI 06 · Deployment & AI Infra Serving and inference optimization API-layer architecture: gateways, streaming, fallbacks Cost and latency engineering Fine-tuning vs. RAG vs. prompting Reliability and operations Project: self-hosted model behind a gateway, load-tested 07 · Emerging Topics Reasoning models and test-time compute Agentic coding and SWE agents Protocols and interoperability Multimodal, realtime, and computer use Small models, distillation, and on-device Project: capability memo on one emerging area Cross-cutting Soft skills — all five sections Hot topics — reviewed and re-dated Capstone — shipped and written up How to Use This Repo Go in order, but don't gold-plate. Modules 01–03 are prerequisites for everything else. 04–06 can be taken in any order once you have them. Build the project before ticking the boxes. Every module ends with one deliverable. Reading about reranking teaches you nothing; watching recall@10 jump from 0.61 to 0.88 teaches you the whole module. Write down what broke. Keep a notes/ folder with the bugs, the wrong assumptions, and the numbers. That log is the actual portfolio artifact. Timebox the reading. Each module lists 2–3 canonical resources per subtopic. Read the primary source once, skim the rest, then go build. Re-read hot-topics.md quarterly. Roughly a third of this material has a shelf life. That file tells you which third. Suggested pace: one module every 1–2 weeks part-time, then 3–4 weeks for the capstone. Faster is fine if you're already shipping AI features at work. Contributions Welcome This is a public learning resource and it gets better with more eyes on it. Useful contributions: Broken or outdated links — the fastest, most valuable fix. Better canonical resources — if a paper or post explains a subtopic more clearly, swap it in and say why. Tool corrections — tooling lists go stale. PRs that add, remove, or re-rank tools based on real production use are especially welcome. New pitfalls — if something cost you a day, it belongs in a Common Pitfalls section. Your project write-ups — link them; concrete examples beat prose. Open an issue to discuss anything structural (new modules, reordering) before writing it. For small fixes, just open a PR. Keep the house style: tight prose, real tool names, links to primary sources over aggregators, no marketing language. Not looking for: SEO filler, links to gated content, "top 10 AI tools" listicles, or vendor pitches. License MIT — fork it, teach from it, rip out the parts you disagree with. Attribution appreciated but the license only asks you to keep the copyright notice. Resources Readme MIT license Activity Stars 1 star Watchers 0 watching Forks 0 forks Report repository