FDE interview questions and how to answer them

FDE interviews test whether you can be trusted with a customer. Every question below appears in real loops; each scores something specific. The interview guide covers the full loop structure.

Question 1: "Our invoicing is a mess. What would you build?"

Scores discovery. The strong answer asks first: who feels the pain, what does today look like step by step, volumes, exception share, systems involved, constraints, and what changes in their week if fixed. Only then sketches a phased approach with assumptions named. Building first scores near zero.

Question 2: "The customer's webhook consumer is down 10 percent of the time."

Scores reliability engineering. Cover retries with backoff and jitter, idempotency keys, a dead-letter queue with a repair runbook, freshness alerting, and a shadow run before cutover. Name what you would measure: delivery rate, duplicate rate, recovery time.

Question 3: "The customer insists on an LLM where rules already hit 94 percent."

Scores AI judgment. Compare: accuracy delta, cost per thousand decisions, latency budget, auditability, failure modes. Propose the honest test: blind evaluation on their data, rules as baseline, model only if it clears the bar with margin. Customers remember who saved them money.

Question 4: "Duplicates appeared two days after your integration shipped."

Scores diagnosis under pressure. Narrate out loud: check idempotency keys, retry configuration, dual-write ordering, clock skew on dedup windows, then quarantine and replay. Communicate early: what users see, what is contained, when the next update lands.

Question 5: "Explain embeddings to our CFO."

Scores executive communication. One analogy, one number, one decision: what it costs, what it buys, when simpler search wins. No jargon without translation.

Worked example: shaping one answer

Weak: "I would use RAG with vectors." Strong: "First I would check who searches what and how often, baseline keyword search on 30 real queries, then test retrieval with an access filter inside the query, and only keep vectors if recall at fixed precision beats the baseline by a margin worth the cost. Here is the scorecard I would publish."

Checklist: answer shape

1. Clarify the real problem in one sentence.
2. Name constraints and unknowns before solutions.
3. Give the phased answer with a smallest first step.
4. State how you would measure and what you would do when wrong.

Related reading

• How to become an FDE in 2027
• FDE vs solutions architect: the real differences
• Writing evaluation scorecards for retrieval systems