Interview prep for applied AI roles at frontier labs
You know the material. Then the interviewer says “walk me through it,” the clock starts, and your best answer comes out half-formed. Interviews aren't purely logical tests: pressure, nerves and a skeptical face in the room decide as much as knowledge does.
So rehearse under pressure. Three role tracks, for Applied AI Engineer, Applied AI Architect and Forward Deployed Engineer: timed drills, graded answers and live, spoken rounds with interviewers and customers who push back.
Start with a 2-minute plan · No installs · No API key · Original questions, never leaked ones
Tell me about a time you slowed down a launch because it wasn't ready. What did you do?
The agent gave wrong refund answers in 4% of tests, so I proposed a 10% canary with an eval gate at 99%…
Why strong engineers stumble
A design prompt with no numbers, a clock and a silent interviewer. Without reps, people jump to a solution and talk until time runs out.
Timed design rounds with a live interviewer who changes the requirements halfway through.
A staff engineer says “I’m not convinced.” A CISO asks you to just confirm. Wanting to be liked makes people agree or overclaim.
Personas trained to push back once, with grading that rewards holding a reasoned position.
Prep sites sell certainty: “they always ask X.” Ten blogs copying one forum post still feel like proof when you’re anxious.
Every hiring claim labelled official, reported or anecdotal, with sources, so you train for what’s real.
Knowing how pressure and persuasion work doesn't only help you read a room. It also protects you from being pushed: by a customer who wants a promise you can't keep, an executive who wants a rounder number, or a prep course that sells fear. Every module trains both.
Pick your role
Start here · Frontier Foundations
How frontier labs hire, the fundamentals every round assumes, the progressive coding screen, reading the labs, and the values round.
Track
Build production LLM systems: prompting, agents, retrieval, evals, latency and cost, then the system design round.
For backend and ML engineers.
Track
Design the customer's solution, run the proof of concept, and defend tradeoffs to executives.
For senior engineers and solutions architects.
Track
Ship working AI inside a customer's messy systems, across ten engagements.
For engineers who like customers.
The loop, practiced
Strip away the company names and candidates describe the same four rounds again and again. Every claim in the course is labelled official, reported or anecdotal, so you know what to trust.
Timed, multi-level Python problems in the progressive format candidates describe: build it, then extend it as the requirements change.
Timed drillsScope an ambiguous product, choose models and retrieval, set an eval gate, and do the cost and latency math out loud.
Graded design answersDefend something you built while an interviewer pushes on your decisions, your failures and your numbers.
Live role-playMission, safety and money questions, answered with true stories and a real reading of what the labs publish.
Mock values interviewHow it works
Written lessons on how frontier labs hire, the fundamentals every round assumes, and the work each role does day to day.
Python in your browser with a simulated Claude SDK, timed coding drills, and customer and executive conversations.
Hidden tests check your code. Claude grades written answers and live sessions against interviewer-style rubrics.
What's inside
Talk to the interviewer out loud with your microphone, hear the replies, and see a timer on every answer. Or type, if you prefer.
The interviewer reveals requirements only when you ask, adds a new constraint mid-round, and grades the diagram you draw as you talk.
Your loop scorecard builds itself: every round's score, your weakest rubric lines, drill times, and the module to fix next.
Set your interview date and hours, take a ten-question self-check, and get a week-by-week plan with a countdown.
Build a project against the real API with your own key, then defend it in the deep-dive rounds and link it from a public record of verified results.
Printable interviewer scripts to practise with a partner, plus five-minute spaced-repetition flashcards for the numbers.
Complete, production-shaped projects to copy into real work: an agent, an integration backend and more, each with offline tests.
Browse them →Sample graded answer
A fictional answer to a design round and the kind of debrief Claude returns: a score per rubric line, a note on each, and the one fix to make first. Your real debriefs are about your own answers.
Sample answer · fictional
E7 · Design a pull-request review assistant (excerpt)
“About 400 PRs a day, median diff around 300 lines. With the touched files and the style guide, a review request is roughly 12K input tokens and 1.5K output. On Sonnet 5.5 at $2 and $10 per million that's about $0.039 a review, or $16 a day.
The bot gets a read-only token and posts comments only; it can never approve or merge. We'll track the thumbs-up rate on its comments and tune the prompt from there. Once it looks good, we turn it on for every repository.”
Sample debrief
69 / 100
Just under the pass mark of 70. Every line gets a score and one specific note.
Good volume and diff-size numbers. Missing: how many comments per PR reviewers will tolerate.
Diff plus touched files plus the repo's style guide is right. No plan for PRs larger than the context budget.
“We'll track thumbs-up rate” isn't a gate. Needs a labelled set of past PRs and a precision target before launch.
Read-only token and no auto-merge: good. Prompt injection through PR text isn't mentioned.
The arithmetic follows, but 1.5K output ignores thinking: Sonnet 5.5 defaults to effort high, so set effort or budget more. Rollout jumps straight to every repo.
Fix first: replace the thumbs-up metric with a release gate on 50 labelled past PRs.
Syllabus
12 modules · 163 lessons · about 72 hours
How Frontier Labs Hire
The loop at Anthropic, OpenAI and Perplexity, round by round: what each round tests, what the AI-use rules are, and why the written application is your first interview.
6 lessons · 1.7h
LLM Fundamentals Interviewers Probe
Tokens, context windows, sampling, tool use, caching, structured outputs, model choice and cost math: the fundamentals every applied round assumes.
8 lessons · 3.4h
Progressive Coding Drills
The multi-level coding screen: design the data model first, extend without rewriting, and stay calm against the clock.
6 lessons · 3.5h
Reading the Labs
How to read and critique what lab leaders publish: steelman, assumptions, strongest counterargument, and what would change your mind.
7 lessons · 2.7h
Values & Behavioral
The round that reportedly fails the most candidates: mission versus money, disagreement, failure and safety tradeoffs, practised live.
10 lessons · 3.5h
How LLMs Work: Model Literacy
Transformers, attention, the KV cache, prefill versus decode, tokenization, sampling, the training pipeline, hallucination, prompt injection, evals and embeddings: the why behind model behaviour that fundamentals and design rounds probe, vendor-neutral where interviews are.
18 lessons · 6.8h
The Modern Agent Stack: MCP, Sub-Agents, Skills and Claude Code
Build and secure an MCP server, fan work out to cheap sub-agents and prove it paid off, package a playbook as an Agent Skill, keep long agents append-only under preserved thinking, red-team an agent, and defend a build-or-buy platform decision to a CTO.
16 lessons · 8.1h
Portfolio Capstone
The Foundations portfolio capstone: build a real project against the Claude API on your own key, in a domain you choose, starting from the capstone-kit repository template. Run four real-model labs, write the README, decision record and eval report, then defend it to a practice reviewer who has read them.
11 lessons · 4.6h
Getting the Job
Career mechanics around the loop: choose the role your evidence supports, decode postings, write resume bullets and outreach notes that survive the screen, handle the recruiter screen, take-homes and the online assessment, and understand levels, equity and negotiation. Every claim is labelled; general hiring practice is marked General.
15 lessons · 5.1h
Agentic System Design
The design-level synthesis of the agent lessons: a method for any 'design an agent' question, workflow patterns versus agents, the agent loop's tools, exits, budgets and loop detection, multi-agent architectures and their token bill, state, memory and context budgets, idempotency, checkpoints and human approval, least privilege and approval tiers, and evaluating and operating agents, for the system design round and for a real customer deployment.
16 lessons · 7.3h
Agent Harness Engineering
Build a coding-agent harness from scratch, the way Claude Code and the Claude Agent SDK are built: the full agent loop with parallel tool calls and every stop reason, a tool registry and runtime, a file edit tool, layered context with compaction and memory, permission modes and hooks, planning and subagents, budgets, checkpoints and resume, tracing and cost, harness evals, defence against injection through tool results, and how to deploy and operate a harness for a customer.
26 lessons · 13.3h
Enterprise Agents End to End
Build one enterprise agent pilot end to end: chatbot vs workflow vs agent, comparing models on a business task, no-code agents in n8n, RAG with citations and the LLM-wiki pattern, scoped tools over APIs and MCP with real authorization, state-graph orchestration (LangGraph) with typed routing and confidence-based escalation, a headless ERP with version-bound human approvals, evaluation and tracing with Langfuse, durable execution with Temporal, and deploying, operating and presenting the pilot. Ends with a complete working solution you can download and run.
24 lessons · 12.3h
12 modules · 114 lessons · about 56 hours
Production Prompting & Structured Outputs
System prompts, JSON schemas, extraction reliability and prompt regression tests.
6 lessons · 2.7h
Tool Use & Agents
Tool loops, MCP servers, sub-agents, failure handling and traces.
8 lessons · 3.6h
Retrieval at Scale
Chunking, hybrid search, reranking, retrieval evals and citations.
6 lessons · 2.7h
Evals as Engineering
Golden sets, LLM-as-judge calibration and regression gates in CI.
6 lessons · 2.8h
Production Concerns
Streaming, latency budgets, prompt caching, rate limits, fallbacks, observability and cost.
6 lessons · 2.4h
LLM System Design Interview
Design rounds for LLM systems: requirements, architecture, evals, failure modes and cost, graded against an interviewer's rubric.
9 lessons · 5.2h
Mock Interview Loop
A full practice loop: timed progressive coding, system design, project deep dive and values.
11 lessons · 6.7h
Agent Evals and Reliability
Evaluate agents the way they fail in production: end-state and trajectory grading in a sandbox, pass^k with clustered intervals, simulated users, error analysis into a ranked taxonomy, overfitting-safe hill-climbing, synthetic and adversarial suites with attack success rate as a gate, rerun policies for flaky cases, and a broken harness to debug. Builds on E4.
15 lessons · 7.9h
Multimodal Input, Streaming and Throughput at Scale
Route PDFs and phone photos into Claude with page-verified extraction and an abstain path, ground answers on the native Citations API, parse the raw streaming protocol with cancellation and TTFT, run async pools and Message Batches jobs that respect rate limits, and find the exact byte that breaks a prompt cache.
12 lessons · 6h
Inference, Embeddings and Model Choices
What happens below the API: prefill and decode, the KV cache, batching and rate-limit pools; embeddings and approximate search; when fine-tuning or distillation is justified; OpenTelemetry for LLM calls; and choosing and migrating between Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1.
12 lessons · 5.5h
TypeScript for Claude
The TypeScript side of the job, run locally with Node (the browser course runs Python only): typed Messages API calls, streaming to a Next.js route handler, tool loops, structured outputs and prompt caching with @anthropic-ai/sdk, and an MCP server with @modelcontextprotocol/sdk. Practised through code-reading checks, a code review, a Python-to-TypeScript port and a local lab.
9 lessons · 3.7h
LLM Inference & Serving in depth
One layer below E10: the inference pipeline and the roofline, KV-cache size math, GQA, paged KV and prefix caching, continuous batching and chunked prefill, speculative decoding and quantization, tensor, pipeline and expert parallelism, the levers you control on a managed API like Claude's, and the latency and cost playbook that decides between self-hosting and the API.
14 lessons · 6.6h
9 modules · 77 lessons · about 32 hours
The Architect's Job
Pre-sales to deployment, solution design docs, and working with Sales and Product.
6 lessons · 2.1h
Enterprise Deployment of Claude
Direct API versus cloud platforms, data handling, security reviews and compliance questions.
6 lessons · 2.3h
Choosing the Approach
Prompting versus RAG versus agents versus fine-tuning; build versus buy; cost models at scale.
6 lessons · 2.4h
Proofs of Concept That Close
Success criteria, bake-offs, eval-driven proofs of concept and honest readouts.
6 lessons · 2.7h
Executive Communication
Architecture one-pagers, objection handling, and role-play with a skeptical CTO and a cautious CISO.
9 lessons · 3h
Reference Architectures
Support automation, document processing, knowledge assistants and coding assistants, each with tradeoffs.
7 lessons · 2.8h
Mock Interview Loop
A full practice loop: case presentation, architecture design, customer role-play and values.
10 lessons · 5.5h
Governance, Commercials and Partners
The conversations that decide enterprise deals after the demo: getting an LLM through a bank's model risk validation, the EU AI Act, NIST AI RMF and ISO/IEC 42001, capacity at peak and how customers buy, fair positioning against an incumbent, honest RFP answers, and three-party work with SIs and cloud providers.
14 lessons · 5.5h
Demos and Enterprise Rollout
Build demos on the customer's own data with an honest scoreboard, recover when they break live, and take the deal further: a Claude Enterprise, Claude Code and connectors rollout, a 90-day adoption plan with works councils in the room, a back-office agent and MCP reference architecture, a public-sector whiteboard round, a startup model-migration consult and a workshop that ends in a build.
13 lessons · 6.1h
13 modules · 151 lessons · about 58 hours
The FDE Role & Mindset
What forward deployed engineers actually do, how the role differs from SWE, solutions engineering and consulting, and the operating loop you'll run on every engagement.
10 lessons · 2.4h
Customer Discovery & Scoping
Turn vague asks into scoped, measurable engagements: discovery calls, stakeholder maps, honest baselines, requirements extracted with Claude, and change requests handled without saying no.
13 lessons · 3.5h
Wrangling Messy Customer Data
Customer data is never clean. Profile it, normalize it, reconcile systems of record, deduplicate entities, label free text with Claude cost-effectively, and report quality so problems get fixed at the source.
14 lessons · 4.4h
APIs & Integrations
Connect to customer systems that paginate, rate-limit, time out and change without warning. Build clients, webhook receivers and sync jobs that survive production, and use Claude to speed up integration work safely.
12 lessons · 3.8h
SQL & Data Modeling on Customer Systems
Query customer databases safely and fluently: answer executive questions, use window functions and cohorts, build a metrics layer, guard production data, and put Claude-powered text-to-SQL behind real guardrails.
12 lessons · 3.8h
Building with Claude: Anthropic SDK Fundamentals
The core of modern FDE work: the Anthropic SDK in depth. Multi-turn state, production prompts with injection defenses, streaming, errors and refusals, prompt caching and real cost, and routing requests across models.
12 lessons · 3.8h
Tool Use, RAG & Agents
Give Claude access to customer systems. Define strict tools, write a traced agent loop, answer from documents with checked citations, gate risky actions behind human approval, and know when not to build an agent.
10 lessons · 3.3h
Evals, Reliability & Cost
Prove your LLM system works before the customer finds out it doesn't. Build eval harnesses, calibrate LLM judges, evaluate agent trajectories and retrieval, gate releases on regressions and noise, and model cost against quality.
12 lessons · 3.9h
Shipping to Production
Deploy into environments you don't control. Map the customer's environment, validate config and protect secrets, log every Claude call safely, alert on what matters, degrade gracefully, debug an incident from logs alone, and hand off with runbooks and a blameless post-mortem.
12 lessons · 3.9h
Capstone: An End-to-End Engagement
Run a full simulated engagement for NorthStar Logistics: scope from discovery notes and a measured baseline, clean and join their messy data, design and build a guarded exception-triage agent, evaluate it and make the ship call, then turn a pilot into fact-checked numbers for the executive readout.
11 lessons · 5.1h
FDE Mock Interview Loop
A full practice loop for forward deployed engineering roles: a take-home customer case, a timed coding drill, a live customer call, and a deep dive on what you built, scored like the real thing.
8 lessons · 5.7h
Enterprise Field Work
Ship inside a locked-down regulated customer: proxies, private networks, SSO, air-gapped installs and platform gaps; data at scale; refusals in regulated domains; writing reusable assets back to product; and the realities of travel, access and juggling customers.
15 lessons · 7h
FDE Drills and the War Room
A higher coding bar for FDE loops: pre-model for Level 4, then four timed progressive drills with interaction-heavy hidden checks (idempotent webhooks, rate-limited sync with backfill, a tool and permission engine, and an AI-allowed SSE client), a 30-minute incident war room, tests graded by mutation testing, a minimal change in a customer's multi-file repo, and a live-coding round where the requirements move.
10 lessons · 7.1h
Who I am
I'm a seasoned AI architect. I worked at Amazon, and today I'm an FDE Architect at an AI startup, working in the Forward Deployed Engineering model and shipping AI-based solutions to enterprise customers.
The interviews for these roles test the job itself: scoping a messy problem, building something that works, proving it with evals, and explaining it to people who decide. Throughout the course you'll find “From my loop” notes with what I've seen work, from Amazon's bar raiser round to a startup's live customer scenario.
I built FDE Playbook to be the prep I wish I'd had.
Pricing
Start free with the first Foundations module, the first lesson of every track and FDE Module 1. Upgrade when you're ready to prepare for real.
For scale: a single one-hour mock with a human interviewer is reported to cost roughly $170 to $340 on popular platforms. This is one payment for every track, unlimited practice and lifetime updates. Human mocks are still worth doing; this is how you arrive ready for them.
Free lessons need no credit card.
Refunds: email us within 14 days (reply to your receipt). A refund ends access.
No. Every question is original, written in the style candidates publicly describe. We never use leaked or confidential material, and every claim about a hiring process is labelled official, reported or anecdotal.
No. It's an independent course. Labs are named only as factual references.
Everyone starts with Foundations. Then pick the role you're applying for: Applied AI Engineer if you build systems, Applied AI Architect if you design and sell them, Forward Deployed Engineer if you ship them inside customers.
No. Everything runs in your browser. Coding exercises use a built-in simulator of the Anthropic SDK that uses the same calls as the real SDK for the features taught; your capstone runs against the real API with your own key.
Yes. The first Foundations module, the first lesson of every track and FDE Module 1 are free, with no credit card.
Yes. Timed drills switch the AI tutor off, the same as real assessments, and the written lessons teach Anthropic's published candidate guidance.
Foundations · C1 · Free
The best feeling in an interview is recognition: you've heard this kind of question, you've been pushed like this, and you know your next move. Start with the free module. It takes under two hours, with no setup and no credit card.
Start free →