FDEInterviews logo
INTERVIEW PREP · FORWARD DEPLOYED ENGINEERS

633+ practice questions from actual FDE loops, with answers that get offers.

Every question ordered easy → hard into a journey per topic. Written by engineers who sat the loops at the labs hiring right now. Browse the full set of Forward Deployed Engineer interview questions by topic and company.

Start practicing free no card · free answers in every topic

Updated · 633 questions added or revised in the last 30 days

Prefer to browse? Follow the map instead: the four stages of preparation, and where to join them given what you already know.

Free diagnosticNo account · 15 min

How ready are you for an FDE interview?

17 scenarios from real loops. Find the gap costing you most, and get a plan wired to real questions and lessons.

Question 1 of 21

Which role are you actually going for?

QUESTION OF THE DAYYour training run isn't crashing, but step time doubled overnight. MFU dropped from 45% to 22%. Triage it.ML Infrastructure & GPUs · hardAnswer →
QUESTIONS MIRROR REAL LOOPS AT
633+PRACTICE QUESTIONS
101COMPANIES COVERED
167CONCEPTS TO MASTER

Grounded in real Forward Deployed Engineer interview loops and written to a senior-engineer editorial bar.

ACCURACY YOU CAN AUDITEvery code sample, SQL query and worked number here is executed and checked before it ships, then re-audited. Find an error we can verify and we add a day of full access to your account.How we check, and what counts →
WHERE DO I START?

The content is the product.

A real, unlocked card from the RAG & Agents journey:
27

Customer's RAG pipeline started hallucinating after a data refresh — debug it live

mediumOpenAI

Resist the urge to touch the prompt. A data refresh changes retrieval, not generation — so measure retrieval before you theorize.

1 · REPRODUCE AND MEASURE
# golden set recall, before touching anything
hits = retriever.search(q, k=20)
print(recall_at_k(hits, gold))  # 0.31 — was 0.86

Recall cratered. The refresh re-ran ingestion but skipped re-embedding, so the index points at stale vectors.

NEW · PRACTICE TESTS

Reading answers is studying. Being tested is knowing.

408 multiple-choice scenarios across all 11 topics, easy to staff-level. Every answer explains itself and links into the worked material, and the questions rotate so a retake is a fresh test.

Take a practice testfree with an account · pick topic, difficulty, length
RAG & Agents · easyexample

A RAG answer is wrong. Which single question splits the failure into its two bug classes?

Was the correct passage present in the retrieved context?
Was the prompt template changed since the last good run?

If the passage never came back, no prompt work will help: fix retrieval. One log lookup separates the two bug classes…

THE LEARNING ENGINE

Not a wall of answers. A system that learns how you learn.

Reading answers from an LLM is passive, and it forgets you the moment you close the tab. This is built on knowledge tracing, the same learning science behind the best intelligent tutoring systems. It quietly models how well you recall each topic and points you to exactly what to study next.

  • One tap after each answer (Got it / Shaky / Review) feeds a live mastery estimate per topic.
  • “Your next focus” sends you to your weakest area, not a random question.
  • Your mastery and journey persist, so every session builds on the last.
MASTERY BY TOPICexample
System DesignStrong 78%
RAG & AgentsStrong 61%
LLM & GenAIDeveloping 44%
MLOpsBuilding 22%
NEXT FOCUSImplement an idempotent agent that survives a mid-run crash
FROM THE ANSWERnot a mock-up
What happens when you call an LLM API TRANSFORMER LAYERS PREDICTION decode: once per output token 1 Your app POST with prompt + params 2 API gateway auth · rate limit · route 3 Tokenize text to token IDs 4 KV cache lookup reuse a matching prefix 5 Embed token ID to a vector 6 Attention tokens read each other 7 Feed-forward network per token, in parallel 8 Logits to softmax a score for every token 9 Sample one token temperature · top-p 10 Stream to the caller until stop token or max Billing and the context window are both counted here, not in words. 1 token ~ 4 chars ~ 0.75 words One vector per token, hundreds to a couple of thousand dimensions. Attention routes information between tokens. The feed-forward network is where the stored knowledge sits. One probability per vocabulary entry, roughly 50k to 200k of them. " Paris" 0.87 " the" 0.04 (illustrative) Steps 3 to 8 run once over the whole prompt in parallel: prefill, which sets time to first token. Every later token pays a full pass of its own: decode, which sets total latency.
DIAGRAMS THAT TEACH

The part a chatbot cannot hand you.

Ask a model what happens when you call an LLM API and you get four paragraphs you have read before. Beside it is what we give you instead: every stage in order, the mechanism printed under each name, and the numbers annotated where you would otherwise have to take our word for it. 160 answers, concepts and lessons now carry one.

  • Drawn for the page it sits in, so it teaches that page rather than the topic in general.
  • Every figure on them is either executed while writing or marked illustrative. Nothing is decorative.
  • Light and dark, readable on a phone, and described in full for a screen reader.
THE CURRICULUM

The only structured curriculum we know of for the FDE job itself.

Certifications teach cloud architecture. AI courses teach models. Neither teaches the embedded, customer-facing half of this job, so we wrote it: 3 courses, 98 lessons, from your first discovery call to deployments behind an air gap.

  • Hands-on labs with runnable harnesses, and worked engagement documents: the scope doc, the pilot proposal, the security packet.
  • Each course ends in a capstone whose artifacts you can bring into an interview.
  • 15 lessons are free, one click away, so you can judge the depth first.
◆ WORK SAMPLES · RUN THEM ON YOUR MACHINE

Reading an answer and producing one are different skills. The loop tests the second.

Each work sample is a small broken thing from a real deployment, handed to you the way a customer hands it over: three symptoms in one ticket, no list of causes. You get the fixtures, a hostile server or a seeded database, and a grader you run locally. Nothing to install, no API key, no account on anything, and every dataset is synthetic and written for the exercise.

The tests going green is the smaller half. Each pack ends with the questions an interviewer asks next, and a reference solution that explains what was being scored rather than only showing the answer. This is the closest thing on the site to the job itself.

Browse the work samples briefs are free to read · the packs come with premium
Cover of The State of Forward Deployed Engineering 2026, an industry report from FDEInterviews.com.
INDUSTRY REPORT

The State of Forward Deployed Engineering 2026

Why 95% of enterprise AI pilots produce nothing measurable, the role the market invented to fix it, and what the 2026 stack, economics and hiring loops actually look like. Sourced, and explicit about which numbers are contested.

Read it free with an account →
Cover of The Forward Deployed Engineer Interview Book by FDEInterviews.com, 2026 edition.
NOVEMBER 2026

The first book on the FDE interview

Not the role in general. The loop: the decomposition round, the practical builds, the applied LLM work, and the customer rounds that decide more offers than anyone admits. Free for premium members on release.

What is inside the book →

One payment unlocks all of it

All 633+ answers across 11 topics, the full concept curriculum, and all 3 courses, for 6 months. One payment, no auto-renewal. Start free, no card needed.

NEW · Premium now also includes The Forward Deployed Engineer Interview Book, free on release in November 2026.

launch price · regularly ₹4,000 / $50

MOST VIEWED QUESTIONS
RECENTLY ADDED
11 TOPICS · CURATED JOURNEYSNew to the role? Read the FDE guide →