Skip to content
The TAI Labs community is now on Skool
TAI Labs
Sign inBook a scoping call

Free self-study book

Cracking AI Interviews at Big Tech

A self-study mini book

A self-study guide to the AI Engineer, Forward Deployed Engineer and AI Product Manager interview loops at Big Tech and the frontier labs, built from October 2026 research, with worked answers, exercises, templates and 4, 8 and 12 week plans.

  • 21 chapters and appendices
  • 23,226 words
  • 3 tracks: engineer, FDE, PM
  • 4, 8 and 12 week study plans

Free, no account, no card. You get the reading link by email and you can unsubscribe from anything else in one click.

Which track are you

Three loops, read three different ways

AI Engineer

LLM system design, hybrid retrieval, agent loops, evals, and coding rounds both classic and AI-assisted. Chapter 7 is the track; chapters 2 to 6 are the ground it stands on.

Forward Deployed Engineer

Decomposing a customer's vague problem on a whiteboard, the customer-simulation round, take-homes, and working in somebody else's environment. Chapter 8.

AI Product Manager

AI product sense, metrics that survive a non-deterministic feature, evals as the spec, unit economics, and Meta-style prototype rounds. Chapter 9.

What is inside

The whole table of contents, before you give us anything

  1. —Start hereWho this is for · How the book works · Reading path by track · How reliable is this material · Diagnostic: pick your starting point · Three rules for every exercise in this book
  2. 1Chapter 1. The 2026 landscape1.1 The three roles in one paragraph each · 1.2 Five shifts that matter for your preparation · 1.3 AI-tool policies by company · 1.4 Market data, with caveats · 1.5 What each loop is really asking · Checkpoint · Exercise (30 minutes)
  3. 2Chapter 2. LLM fundamentals you must explain cold2.1 Tokens and tokenisation · 2.2 Context windows · 2.3 Attention and the transformer, at interview depth · 2.4 Embeddings · 2.5 How chat models are trained · 2.6 Sampling controls · 2.7 Structured output and tool calling · 2.8 Hallucination · 2.9 Prompting, RAG and fine-tuning: the commit-to-an-answer decision · 2.10 Worked example: estimate a feature's cost · Checkpoint · Exercise (45 minutes)
  4. 3Chapter 3. RAG from demo to production3.1 The pipeline, stage by stage · 3.2 Where RAG fails: a diagnostic table · 3.3 Hybrid retrieval in code · 3.4 Evaluating RAG: measure the two halves separately · 3.5 Worked design: permissioned enterprise assistant over 10 million documents · 3.6 Worked debugging: "it hallucinates even though the right document exists" · 3.6.1 Advanced topics worth one paragraph each · Checkpoint · Exercise (90 minutes)
  5. 4Chapter 4. Agents, tools and MCP4.1 What an agent is · 4.2 Workflows versus agents · 4.3 Error compounding, with numbers · 4.4 Tool design · 4.5 Memory and state · 4.6 Safety: prompt injection and the lethal combination · 4.7 Evaluating agents · 4.8 The Model Context Protocol (MCP) · 4.9 Worked design: an expense-report agent · Checkpoint · Exercise (2 hours)
  6. 5Chapter 5. Evals, the senior signal5.1 What an eval is · 5.2 The workflow, in order · 5.3 Grader types and when to use each · 5.4 Pitfalls interviewers probe · 5.5 A minimal eval harness · 5.6 Worked example: a support-ticket classifier · 5.7 How each track uses evals · Checkpoint · Exercise (2 hours)
  7. 6Chapter 6. Cost, latency and safety6.1 Measure before you optimise · 6.2 Cost and latency levers · 6.3 Worked answer: "cut cost and latency by 50 percent without hurting quality" · 6.4 Unit economics for PMs and FDEs · 6.5 Safety and responsible deployment · 6.6 Safety questions you should be ready for · Checkpoint · Exercise (90 minutes)
  8. 7Chapter 7. Track 1: AI Engineering7.1 Loop anatomy · 7.2 Topics by seniority · 7.3 Coding rounds, classic and AI-assisted · 7.4 The LLM system design framework · 7.5 Second worked design: a support-agent reply copilot · 7.6 ML fundamentals still tested at Big Tech · 7.7 Project deep-dive: how to present your best project · 7.8 What differentiates strong candidates, and common mistakes · Checkpoint · Exercise (3 hours)
  9. 8Chapter 8. Track 2: AI Forward Deployed Engineer8.1 Loop anatomy (composite of reported loops) · 8.2 The decomposition framework · 8.3 The customer-simulation round · 8.4 Discovery question bank · 8.5 Take-home playbook · 8.6 Question bank with strong-answer outlines · 8.7 Working in someone else's environment · 8.8 Common mistakes · Checkpoint · Exercise (3 hours)
  10. 9Chapter 9. Track 3: AI Product Management9.1 Loop anatomy by company (mostly candidate-reported) · 9.2 The AI product sense framework · 9.3 Worked answer: "Users say our AI is confident but wrong. Fix it." · 9.4 Metrics for AI features · 9.5 Evals as the product spec · 9.6 The vibe-coding drill (for Meta-style prototype rounds) · 9.7 Technical concepts AI PMs must know cold · 9.8 Pricing and unit economics: a worked example · 9.9 Behavioural rounds for AI PMs · 9.10 Common mistakes · Checkpoint · Exercise (3 hours)
  11. 10Chapter 10. Company playbooks10.1 OpenAI · 10.2 Anthropic · 10.3 Google and DeepMind · 10.4 Meta · 10.5 Microsoft · 10.6 Amazon · 10.7 NVIDIA · 10.8 Apple · 10.9 Palantir · 10.10 The universal last-48-hours checklist · Checkpoint · Exercise (1 hour)
  12. 11Chapter 11. Behavioural mastery11.1 STAR-L: a better frame than STAR · 11.2 Build your story bank · 11.3 Mapping stories to company values · 11.4 A worked story (illustrative numbers) · 11.5 Mission and safety answers for the labs · 11.6 Delivery rules · 11.7 Twenty prompts to practise with · Checkpoint · Exercise (3 hours across a week)
  13. 12Chapter 12. Study plans12.1 A weekly rhythm that works · 12.2 The 4-week plan (about 12 to 15 hours a week) · 12.3 The 8-week plan (about 10 to 12 hours a week) · 12.4 The 12-week plan (about 8 to 10 hours a week) · 12.5 Mock interviews: how to get value · 12.6 When you are interviewing at several companies · Checkpoint · Exercise (30 minutes)
  14. 13Chapter 13. Portfolio projects that signal strength13.1 What every strong project has · 13.2 Project 1: a RAG system with a public eval report · 13.3 Project 2: a tool-using agent with guardrails · 13.4 Project 3: an MCP server for a real data source · 13.5 Project 4: an FDE-style case study · 13.6 Project 5: a PM artifact pack · 13.7 Publishing and presenting your project · 13.8 A README template · Checkpoint · Exercise (30 minutes)
  15. 14Chapter 14. Applications, referrals and offers14.1 The resume · 14.2 Writing your own application · 14.3 LinkedIn and visible proof · 14.4 Referrals and outreach · 14.5 Process hygiene · 14.6 Compensation data: how to read it · 14.7 Negotiation playbook · 14.8 Offer comparison sheet · Checkpoint · Exercise (2 hours)
  16. 15Chapter 15. Day-of execution15.1 The setup checklist · 15.2 The four narration habits · 15.3 Time boxes · 15.4 Rounds with an AI assistant · 15.5 When you freeze or go blank · 15.6 Asking good questions at the end · 15.7 After each round · Checkpoint · Exercise (1 hour)
  17. 16Chapter 16. Self-test: rubrics and answer keys16.1 The universal scoring rubric · 16.2 Rapid-fire concept check (answer in under 30 seconds each) · 16.3 AI Engineer scenarios · 16.4 FDE scenarios · 16.5 AI PM scenarios · 16.6 Behavioural self-check · 16.7 Your readiness dashboard · Exercise (2 hours)
  18. 17Appendix A. Templates and checklistsA.1 LLM system design canvas · A.2 Decomposition worksheet (FDE) · A.3 AI product sense canvas (PM) · A.4 Eval plan template · A.5 Recruiter email: AI-tool rules per round · A.6 Thank-you email after a round · A.7 Mock interview scorecard · A.8 Error log · A.9 Company tracker · A.10 One-page pre-interview checklist
  19. 18Appendix B. Glossary
  20. 19Appendix C. Reading and practice listC.1 Books and long-form · C.2 Essays and articles · C.3 Company materials · C.4 Courses and practice · C.5 Practice habits that pay off
  21. 20Appendix D. Sources and confidence notesD.1 How firm is each kind of claim · D.2 Known gaps · D.3 Sources consulted by the research pass · D.4 A closing note

A page from it

Chapter 3, on why a working RAG demo still fails

3.2 Where RAG fails: a diagnostic table

SymptomLikely stageFix to try first
Right document exists, answer is wrong or missingRetrievalCheck whether the chunk was retrieved at all. Try hybrid search, query rewriting, better chunking.
Right chunk retrieved but ranked lowRankingAdd or tune a reranker, increase candidate count.
Right chunk in the prompt, model ignores or misreads itGenerationReorder context, tighten instructions, reduce noise, shorten context.
Answers cite the wrong version of a policyIngestion and metadataAdd effective dates, version filters, deduplication.
Users see documents they should notPermissionsFilter at retrieval time by the user's access. Never rely on the model to withhold.
Tables and figures answered badlyParsingUse layout-aware parsing, store tables as structured text.
Answers sound confident without sourcesPrompt and verificationRequire citations, add a groundedness check, allow refusal.
Quality drifts after a content updateIndex freshnessAdd re-indexing jobs, and run the eval suite after every change.

The book is the self-study path

If you would rather do it in a room with other people, weekly reviews and someone reading your work, that is what the cohorts are. Same material, more pressure.

See the courses