1,801 interview questions with answer frameworks for AI/ML engineering, system design and architecture loops.
Most interview banks are lists of questions. This one pairs every question with an answer framework — the mechanism, the tradeoff, the follow-up the interviewer asks next, and the trap — because knowing a definition rarely survives the second question.
It covers the full modern surface: classic ML and statistics through transformers, RAG, agents, MCP/A2A interoperability, serving and inference optimisation, LLMOps, evaluation, safety, governance, cloud agent deployment, and the leadership and communication rounds that decide senior offers.
| 📋 Question bank | 1,801 questions across 58 sections, difficulty-tagged (⭐ Standard, ⭐⭐ Hard, ⭐⭐⭐ Principal). Hover or tap any question on the site to reveal its answer. |
| ✅ Answer frameworks | An answer for every question — not a definition, but what a strong candidate actually says. |
| 🎯 Role-based map | Pick your target role (Staff MLE, Principal AI Lead, GenAI Engineer, Platform, MLOps, Research, Security, Solutions Architect, Data Scientist, EM) and see exactly which sections to complete, with progress tracking. |
| 🎙️ Live Interview | An agentic interviewer: asks aloud, listens, then follows up on what you actually said — adaptive with an API key, gap-based without one. Ends with a scored summary. |
| 🎤 Mock simulator | Voice-enabled. Questions read aloud, answer out loud with live transcription, then get scored — local concept-coverage analysis built in, or LLM feedback with a free API key. |
| 🏗️ Diagrams + Patterns | 69 Mermaid architecture diagrams and conceptual patterns, each with flow, worked example and real-world industry usage. |
| 💻 Code solutions | Runnable implementations for the coding round. |
| 🗺️ Study paths | Day-by-day 2-week plans for three role tracks. |
| ⚡ Cheat sheet · Glossary | Day-of revision and 60+ acronyms. |
| 🏢 Company prep | Reported loop structures for Anthropic, OpenAI, Google DeepMind and enterprise tracks. |
| 📚 Sources | Cited industry sources and honest notes on what has been fact-checked and what hasn't. |
You are not meant to read 1,801 questions.
- Pick your role — it narrows 58 sections down to the ones your loop actually tests, split into Core / Important / Optional.
- Work the Core sections, reading questions and revealing answers as you go.
- Drill in the simulator — answer out loud and self-grade. Articulation under pressure is the thing being tested, not recognition.
- The week before: cheat sheet, multi-turn drills (Section 51), and company prep.
Beyond standard recall and design questions, the bank includes formats that senior loops actually use to separate candidates:
- Multi-turn interviewer drills (Section 51) — full transcripts where the interviewer pushes back turn after turn as each hypothesis is eliminated.
- Production incident triage (Section 50) — "429s appeared in production, what do you check, and what does each signal rule out?"
- Spot the flaw (Section 52) — plausible-looking designs and code with real defects to find.
- Estimation and capacity arithmetic (Section 53) — KV cache sizing, GPU counts, cost per task, latency budgets.
- Executive communication (Section 54) — explaining to a board, a CFO, a regulator, a security team.
The bank is available as JSON for building your own tools:
data/questions.json— every question with id, section and textdata/answers.json— answer lookup keyed by question id
All 58 sections
| # | Section | Questions | Range |
|---|---|---|---|
| 1 | Strategy, Vision & Technical Leadership | 25 | 1–25 |
| 2 | Leadership & Behavioral | 40 | 26–65 |
| 3 | Classic ML Fundamentals | 70 | 66–135 |
| 4 | Statistics & Probability | 35 | 136–170 |
| 5 | Deep Learning Fundamentals | 55 | 171–225 |
| 6 | Computer Vision | 30 | 226–255 |
| 7 | NLP Fundamentals (Pre-LLM) | 30 | 256–285 |
| 8 | LLM & Transformer Fundamentals | 60 | 286–345 |
| 9 | Prompt Engineering & Structured Outputs | 30 | 346–375 |
| 10 | RAG & Retrieval | 45 | 376–420 |
| 11 | Vector Databases & Embeddings | 30 | 421–450 |
| 12 | Agentic AI & Multi-Agent Systems | 45 | 451–495 |
| 13 | LLM System Design / GenAI Architecture | 60 | 496–555 |
| 14 | Classic ML System Design | 45 | 556–600 |
| 15 | Model Serving & Inference Optimization | 45 | 601–645 |
| 16 | LLMOps & MLOps | 55 | 646–700 |
| 17 | Feature Stores & Feature Engineering | 25 | 701–725 |
| 18 | Data Engineering for AI | 40 | 726–765 |
| 19 | Cloud ML Platforms | 30 | 766–795 |
| 20 | DevOps & Infrastructure for AI | 35 | 796–830 |
| 21 | LLM Evaluation | 35 | 831–865 |
| 22 | Safety, Guardrails & LLM Security | 40 | 866–905 |
| 23 | Governance, Ethics & Responsible AI | 35 | 906–940 |
| 24 | Time Series & Forecasting | 20 | 941–960 |
| 25 | Recommender Systems | 20 | 961–980 |
| 26 | Coding & Algorithms for ML | 25 | 981–1005 |
| 27 | Open-Ended Architecture Design Prompts | 30 | 1006–1035 |
| 28 | Rapid-Fire Depth Probes | 55 | 1036–1090 |
| 29 | Enterprise AI Governance, Frameworks, Platforms & Executive Communication | 50 | 1091–1140 |
| 30 | Enterprise Agent Interoperability (MCP, A2A) & Advanced RAG | 40 | 1141–1180 |
| 31 | Cloud-Native Agent Deployment: AWS, Azure, GCP | 26 | 1181–1206 |
| 32 | Multimodal AI & Vision-Language Models | 30 | 1207–1236 |
| 33 | Fine-Tuning, Adaptation & Model Compression | 30 | 1237–1266 |
| 34 | Responsible AI: Fairness, Bias & Explainability | 25 | 1267–1291 |
| 35 | Generative AI: Image, Video & Code Generation | 25 | 1292–1316 |
| 36 | Speech, Audio & Conversational AI | 20 | 1317–1336 |
| 37 | Edge AI, On-Device ML & Federated Learning | 20 | 1337–1356 |
| 38 | Distributed Training & Large-Scale ML Infrastructure | 20 | 1357–1376 |
| 39 | Data-Centric AI, Labeling & Synthetic Data | 20 | 1377–1396 |
| 40 | Search, Ranking & Information Retrieval | 20 | 1397–1416 |
| 41 | Causal Inference & Experimentation | 15 | 1417–1431 |
| 42 | Graph ML & Knowledge Graphs | 15 | 1432–1446 |
| 43 | Advanced Agentic Systems, Tool Use & Multi-Agent Frameworks | 25 | 1447–1471 |
| 44 | AI Hardware Acceleration, Low-Level Kernels & Compute Engineering | 25 | 1472–1496 |
| 45 | AI Security, Red Teaming, Adversarial ML & Guardrails | 25 | 1497–1521 |
| 46 | Long-Context Mechanics, State Space Models (SSMs) & KV-Cache Optimizations | 25 | 1522–1546 |
| 47 | Domain-Specific AI Architecture (Robotics, Bio, Finance & Software Agents) | 25 | 1547–1571 |
| 48 | Deep-Dive Agentic Frameworks, AI Gateway Architecture & Token Budget Engineering | 25 | 1572–1596 |
| 49 | Enterprise Cloud AI Deployment Architectures (AWS, Azure & GCP) | 25 | 1597–1621 |
| 50 | Production Incident Triage & Live Debugging | 25 | 1622–1646 |
| 51 | Multi-Turn Interviewer Drills | 15 | 1647–1661 |
| 52 | Spot the Flaw: Design & Code Critique | 20 | 1662–1681 |
| 53 | Estimation, Capacity & Cost Arithmetic | 20 | 1682–1701 |
| 54 | Executive & Stakeholder Communication | 15 | 1702–1716 |
| 55 | Voice, Vision & Computer-Use Agents | 25 | 1717–1741 |
| 56 | Agent Memory & Context Engineering | 20 | 1742–1761 |
| 57 | Conformal Prediction & Uncertainty Quantification | 20 | 1762–1781 |
| 58 | Optimisation & Operations Research for AI Systems | 20 | 1782–1801 |
- Most answers have not been independently fact-checked. One research pass verified roughly 30 answers against current sources; the rest are written from model knowledge.
sources.mddocuments what was checked and what changed. Verify anything you plan to state as fact. - Fast-moving areas go stale. Serving frameworks, agent protocols and regulatory deadlines shift within months. Treat 2025–2026 specifics as needing a re-check.
- Depth is uneven. Recently written sections are substantially deeper than some older ones; this is being addressed section by section.
- Company prep is community-reported, not official, and interview processes change.
git clone https://github.com/krunlp/AI-Prep-Buddy.git
cd AI-Prep-Buddy
./scripts/serve.sh # → http://localhost:8000 (no dependencies)That serves index.html, interview.html, simulator.html and roles.html — everything interactive. For the markdown pages too:
./scripts/serve.sh jekyll # → http://localhost:4000/AI-Prep-Buddy/A server is required — the interview and simulator pages fetch data/*.json, and browsers block fetch() over file://, so opening the files directly leaves them blank. localhost is a secure context, so the microphone works without HTTPS.
On macOS, don't use the system Ruby at /usr/bin/ruby (2.6.x — too old for Jekyll and needs sudo). The script detects it and tells you what to do; brew install ruby is the fix.
Corrections are especially welcome — a wrong answer in an interview bank is worse than a missing one. See CONTRIBUTING.md for the source-of-truth workflow and the integrity checks that run in CI.