ai-engineering-from-scratch

Practise for AI engineering interviews against the course itself. This edition adds an interview bank of 482 questions and 60 measured incident scenarios, a mock interview skill, and code blocks rendered at build time.

Interviews · Scenarios · Skills · Start learning

Interview bank

• 482 questions in 23 cards, from math and classical ML to inference serving, agents, protocols, sandboxes and system design.
• Each question shows the prompt first and folds the answer, so you answer before you read. A short answer, a full answer, follow-ups from L1 to L4, and the wrong answers that interviewers hear.
• Every answer links to the lesson or manual section that teaches it. Every number in an answer comes from a file in the repository.
• 7 role paths pick the cards for a role, such as inference engineer or applied AI engineer.

Incident scenarios

• 60 scenarios built from real runs on the L4 and the GB10 in the SGLang, vLLM and Miles manuals.
• Each one gives the situation and the evidence, step by step: logs, metrics and configs, byte for byte from the run. Then the walkthrough, the resolution, and the wrong turns people take.

Mock interviews with an agent

The new interview-prep skill runs a mock interview in your coding agent.

• It asks one question or one scenario and waits for your answer. It never shows part of the answer first.
• It grades your answer against the bank, asks follow-ups one at a time, and sends each gap to the lesson that teaches it.
• It keeps a record of your sessions in INTERVIEW-LEARNING.md.

Site

• Code blocks on practice pages, cheat sheets and manuals are rendered once at build time with Shiki. Terminal blocks copy only the commands, and output blocks have no Copy button.
• New Miles RL cheat sheet.

Fixes

• Review fixes across the interview answers, checked against their lessons: optimizer and schedule edge cases, attention and tokenizer code, cache pricing, eval statistics, and protocol and sandbox details.
• The generated interview indexes are now built on main after each merge, so contributions to the bank no longer conflict with each other.

Full changelog: https://github.com/rohitg00/ai-engineering-from-scratch/compare/v2026.10.5...v2026.11.1

Learn beside your terminal, read two new hardware manuals, and practise every lesson. This edition adds the lesson screen for agent tutors, vLLM 101 and Miles 101, 47 practice tasks, and two new inside-the-machine pages.

Start learning · Skills · Manuals · Practice

Lesson screen for agent tutors

The new lesson-screen skill opens a browser tab next to the terminal while a tutor teaches.

• It shows the lesson's own interactive figures, code traces with values from a real run, and clickable predictions.
• The tutor can build a figure for the lesson while it teaches: a blueprint, a mechanism, an isometric machine, or a 3D model. It checks each build with screenshots before it shows it.
• After the tutor's quiz, a practice page gives a brief with key-term cards, a timed drill, spaced review at 1, 3, 7, and 16 days, and links to go deeper.
• Clicks are hints for the tutor. They never count as quiz answers or evidence. The certification tutors use the screen in Lesson mode only.

All six lesson tutors offer it. Without a local browser, they give the lesson page instead.

Manuals

• vLLM 101 and Miles 101 are published, next to SGLang 101, with measured, clock-locked GPU results and hardware-specific cover plates.
• SGLang 101 has 168 audited claims checked and corrected, research caveats, and new sweep configs.
• Manuals use their own words instead of quotes, pin evidence links to the built commit, and are easier for search engines and AI agents to find.

Practice and machines

• 47 practice tasks across implement, research, train, and lab kinds.
• Inside a Decompiler and Inside an LLM serving GPU machines, graded reverse-engineering labs, and step-synced lesson machines.

Site

• The navbar has a System, Light, and Dark theme menu in place of the old N/D switch. System follows your OS live.
• The catalog shows lesson numbers and done checkboxes.
• Quizzes are placed once per lesson. Highlights and read-aloud controls no longer overlap on phones and tablets.

Fixes

• Source errors corrected across phases 00 to 19, with phase 12 cheat sheets synced.
• Apple Silicon (MPS) device support across lessons. Thanks @Jayantkhandebharad.
• Stale translations repaired, and reviewed translations protected.
• Tutor skill lesson counts match the catalog.

Full changelog: https://github.com/rohitg00/ai-engineering-from-scratch/compare/v2026.10.4...v2026.10.5

Highlight what matters, keep notes beside the text, and carry a compact reference into your next build. This edition adds a reading workspace, 154 cheat sheets, and the ready edition of SGLang 101.

Start learning · Cheat sheets · SGLang 101

Highlights and notes

Select text in a lesson or manual, choose one of four colors, and add a note. Open My Highlights to search, filter, jump back to a passage, or manage several highlights together.

• Export Markdown for your notes or JSON for backup and import.
• Move highlights between devices with a sync file or share link.
• Review uncertain matches when a passage changes.
• Use the same tools on mobile, in light and dark themes, and with forced colors.

Highlights stay in your browser. Use an export or share link to move them elsewhere.

154 cheat sheets

Find commands, formulas, diagrams, defaults, and troubleshooting notes across all 542 lessons, four manuals, and 78 projects. Browse 12 tracks or search by topic, task, command, or term. Follow a source link when you need the full explanation.

The complete collection is attached as aiefs-cheatsheets.pdf.

SGLang 101 is ready

Serve a model, size its KV cache, measure the concurrency knee, and decide when to split work across GPUs. The manual covers SGLang 0.5.22 in 28 sections plus six reference sections, with 63 figures and a cover plate.

The examples distinguish measured results from code defaults and reported claims. This edition includes Apple Silicon, GB10, cooled L4, and Miles rollout evidence, including CUDA graphs, attention backends, and three training seeds.

Read it online or download aiefs-manual-sglang-101.pdf below.

More in this edition

• A beginner guide and 142 exercise hints across 34 lessons in Phases 0 and 1.
• Project selection by role and lesson search inside phase dialogs.
• Reverse Engineering 101 updated for rea 6.7.0.
• Improved article navigation on smaller screens and a clean roadmap URL.
• Fixes for links containing parentheses, PDF math rendering, and deployment bundles.

Downloads

All six English books are included in PDF and EPUB. The four manual PDFs cover A2A, Docker Sandboxes, Reverse Engineering, and SGLang. The complete cheat-sheet PDF and six preview images are also attached.

Thanks to our contributors

Thank you to @Abhishek-1337 for lesson search in phase dialogs, @AkshayKathiriya and @dajiaohuang for fixes to links containing parentheses, and @ALANDVO for adding lesson summary cheat-sheet support. Thanks also to everyone who reported problems, reviewed changes, and helped improve the course.

[Full changelog](https://github.com/rohitg00/ai-engineering-from-scratch/compare/v2026.10.3...v2026

Edition 2026.10.3 adds a reverse engineering track, a third manual, and LaTeX math in the lessons. The curriculum grows from 523 to 542 lessons across 20 phases.

Learn reverse engineering

Fourteen new Phase 13 lessons teach how to read a compiled program you built yourself, from file bytes to evidence-backed pseudocode (#597).

• The lessons cover program anatomy (ELF, Mach-O, PE), disassembly and control-flow graphs, lifting to IR and SSA, calling conventions on System V, Windows x64 and AArch64, compiler idioms, linking and stripped binaries, data and type recovery, dynamic analysis, bytecode formats (Python, JVM, .NET, WebAssembly), binary diffing and symbolic execution, and a ground-truth scorecard for AI-assisted analysis.
• Each lesson runs real tools on its own fixtures: llvm-objdump, Ghidra, radare2, rizin, angr, lldb, LIEF, javap and wasm-objdump, among others.
• Two graded projects: binary-inspector (6 stages, 91 tests) and mini-decompiler (8 stages, 105 tests), from an ELF loader to structured C that recompiles and compares equal.
• A focused learning path and the learn-reverse-engineering tutor skill take a learner through the 14 lessons in order.

Read Reverse Engineering 101

Reverse Engineering 101 teaches rea (rea-agents 6.5.0) with the Ghidra 12.1.4 provider, from the first rea doctor to an agent that keeps an evidence ledger over MCP. It has 26 sections in 8 parts, six reference sections, and 37 figures. The PDF is attached below as aiefs-manual-reverse-engineering-101.pdf (#601).

• Every listing comes from a recorded run of the real CLI and MCP server. capture/run.py --check captures again and compares 65 files.
• Every target is a fixture program written for the manual.

Read the math

• Lesson formulas render with LaTeX through KaTeX on the site, and as math on GitHub (#598). Formulas use only macros that GitHub allows, and CI checks that (#599).

New lessons and projects

• Five lessons fill gaps in the curriculum, and broken links are repaired (#584).
• Half the project roadmap is complete, with two new validated projects (#565).
• The architecture calculator in lesson 10/14 reads head_dim, tied heads, and sliding-window layers (#511).

Site

• Lessons and their translations are rendered on the server, so crawlers and readers without JavaScript get the full text (#570).
• Glossary terms have their own pages, and each phase has a hub page (#574).
• Every page has its own social card (#576) and a crawlable favicon (#581).
• Inside an AI Agent is an interactive machine that shows each step of an agent run (#590).
• A /skills page lists the course skills, with five new figure and manual skills (#593).
• On wide screens, the lesson breadcrumb stays below the lesson (#600).

Contributors

Thank you to everyone who contributed to this edition:

• @FenjuFu fixed the architecture calculator in lesson 10/14 (#511)

Sponsors

Thank you to SerpApi for supporting the curriculum. SerpApi provides a Web Search API for AI apps, available in Markdown and JSON.

Books and manuals

All six English volumes are rebuilt as EPUB and PDF files from this edition's source and attached below, with the A2A 101, Docker Sandboxes and Docker Agent 101, and Reverse Engineering 101 manuals beside them. Interactive figures, graded quizzes, and runnable labs remain accessible through each chapter's online links. The website and README book links continue to use the latest release assets.

Full changes since Edition 2026.10.2