guocong-bincai/ai-interview-guide

ai-interview-guide is a free, open-source interview-preparation collection for people applying to AI-related engineering jobs. It is organized as a question bank plus study materials, aimed at covering many different AI roles — not just one job title. The README describes it as a "systematic interview guide for all AI positions," with role-specific study routes so you don't have to work through everything from start to finish.

What's inside

• 28 topic areas and 720 questions. Topics include LLM basics, prompt engineering, RAG systems, Transformer architecture, AI agents, vector search, model training, inference optimization, AI safety, production deployment, multimodal AI, and more. (One section says "673 interview questions" while another says "720 questions across 28 topics" — both numbers appear in the README.)
• A diagram for every question. The project pairs each question with an illustrated "knowledge diagram" in an anime/comic style. These drawings are not decoration: each one is made after reading the full question and answer, and is meant to capture the core mechanism, the engineering flow, common mistakes, and a one-line memory hook. The idea is to memorize the picture first, then expand into a spoken answer.
• 11 slide decks (495 pages total) called "Dive into LLMs." These are course-style PDFs on topics like fine-tuning and deployment, prompting and chain-of-thought, knowledge editing, math reasoning, text watermarking, jailbreak attacks/defense, text steganography, multimodal models, GUI agents, agent safety, and RLHF alignment.

How the answers are designed

The project's answers deliberately go beyond "what is X." They are built to also explain: why this choice was made, how to verify it, and how to weigh quality, latency, cost, safety, complexity, and alternatives. The README calls this the "engineering trade-off" emphasis.

Each question is meant to be practiced in three versions:

1. 30-second version — give the conclusion and core judgment first.
2. 3-minute version — add principles, option comparisons, and engineering trade-offs.
3. Follow-up version — explain how to verify, what to do when it fails, and why not use an alternative approach.

Role-specific learning paths

The README suggests a different reading order for each role, for example:

• LLM Application Engineer: LLM basics → Prompt → RAG → Agent basics → Production deployment → AI system design
• RAG Engineer: LLM basics → RAG → Vector index → Advanced RAG → Safety/evaluation → Agent observability
• Agent Engineer: Prompt → Agent basics → Planning/reflection → Multi-agent → MCP/Skills → Observability
• Model Training/Inference Engineer: Transformer → Model training → Inference optimization → Inference frameworks → Production deployment
• Forward Deployed Engineer (FDE): Project experience → Python engineering → AI system design → Production deployment → FDE topic
• Multimodal Engineer: Transformer → Multimodal → Multimodal agents → Inference optimization → Production deployment
• AI Safety/Evaluation Engineer: LLM basics → Prompt → Safety/evaluation → Observability → Production deployment

Content quality rules

The project states it prioritizes "question quality over pure quantity." Specific rules from the README:

• No unsourced precise performance numbers.
• No fabricated "big company interview" labels.
• No first-person project achievements.
• Before submitting changes, contributors must manually check for broken links, in-page anchors, question numbers, duplicate titles, and sources.
• An automated audit checks for broken links, duplicate questions, and entries missing required data.

How to use it

The README's quick-start advice:

• First-time learner: follow the role route above.
• Quick review before an interview: start from the "interview questions by company" section, then go back to weak topics.
• Preparing for deep project questions: practice the project-experience and AI-system-design sections.
• Preparing for an FDE role: practice the FDE section (open-ended problem breakdown, customer delivery, production AI systems).
• General advice: lead with your judgment, then explain the knowledge; replace placeholder example numbers with real data from your own experiments or projects; state when a solution works and when it fails, plus how to monitor and roll back; and verify time-sensitive details (product versions, paper results, performance claims) against official sources.

Additional details

• The README is translated into Chinese, English, Japanese, Korean, Brazilian Portuguese, and Spanish.
• The project is licensed under MIT.
• The maintainer accepts voluntary "tips" (via WeChat/Alipay QR codes) but states this is completely optional and unlocks no extra content; starring the repo, sharing it, or submitting real interview feedback is described as equally valuable support.
• The repo includes two governance documents: a content-quality specification (CONTENT_QUALITY.md) and an illustration-production specification (ILLUSTRATION_GUIDE.md).