Interview prep
Crack the FDE interview loop
What each round tests, what separates a senior answer, how interviewers score you, and exactly how to practise. Loops vary by company, but these 8 rounds cover nearly all of them.
- 1
Recruiter & role-fit screen
30 min · conversation
Whether you understand the FDE role, can tell a clear story, and are comfortable in front of customers.
Senior signal: A crisp story of things you delivered for customers, with numbers attached.
Skill guide: Ownership & business valueHow to prepare
- A two-minute "why FDE" answer.
- A one-line contrast: SWE makes it correct, SE shows it fits, FDE makes it work for one customer and proves value.
- Three deployment stories, each with a measurable result.
- 2
Practical coding
45–60 min · often a shared doc, no IDE
Clean, correct code on small real-world problems: parsing, reconciliation, pagination, rate limits, caching.
Senior signal: States invariants up front, traces tests by hand, and evolves the design for scale without breaking its contract.
Skill guide: Practical coding & dataHow to prepare
- Drill ~60 problems across the pattern families in the coding guide.
- Practise narrating while you code. Silence costs points.
- Keep a log of one mistake per session and re-drill it.
- 3
Debugging & codebase lab
60 min · unfamiliar code
Reading code you didn't write, reproducing bugs with tests, and making minimal, correct fixes.
Senior signal: Writes the failing test before the fix, ranks hypotheses, and rejects plausible-but-wrong AI-suggested patches.
Skill guide: Practical coding & dataHow to prepare
- Practise on small buggy modules: caches, retry loops, paginators.
- Always reproduce first, then fix, then keep the test.
- Inject clocks instead of using the wall clock in tests.
- 4
Decomposition / customer scenario
45 min · information revealed in stages
Turning a vague mission into one valuable workflow, an MVP, success metrics and a rollout plan.
Senior signal: Asks questions whose answers change the design, finds the real bottleneck, and says clearly what not to build.
Skill guide: Problem discovery & reframingHow to prepare
- Run 20 cases at 45 minutes: 5 read, 10 clarify, 10 design, 10 MVP/evals/rollout, 5 summary.
- Practise modelling the core objects and their states out loud.
- After each new fact, say what changed in your plan.
- 5
Integration & GenAI system design
45–60 min · whiteboard
Designing inside a customer's systems: data flow, permissions, failure handling, evals and rollout.
Senior signal: Chooses the simplest reliable design, enforces permissions outside the model, and gives capacity estimates with stated assumptions.
Skill guide: Applied GenAI: RAG, agents & toolsHow to prepare
- Use one skeleton: goals → constraints → architecture → data flow → safety → evals → scale → operations → trade-offs.
- Rehearse six integrations aloud: CRM, ticketing, SSO, warehouse, document store, ERP.
- Mark clearly where the model reasons and where code controls.
- 6
Client role-play
30 min · interviewer plays the customer
Staying calm with a pushy exec, a blocking security team, an angry user or a failed demo.
Senior signal: Acknowledges the goal, offers a bounded alternative, documents the risk, and leaves the decision with the customer: no caving, no arguing.
Skill guide: Stakeholder communicationHow to prepare
- Record yourself on eight scenarios: deadline pressure, security block, low adoption, distrust of AI, broken integration, failed demo, scope creep, cost overrun.
- Use the trade-off sentence: this choice, this benefit, this trade-off, this validation.
- 7
Production incident
30–45 min · live troubleshooting
Triage across data, retrieval, prompts, tools, permissions and infrastructure, plus customer communication.
Senior signal: Contains first, preserves evidence, describes impact in business terms, and closes with a postmortem that has owners.
Skill guide: Reliability & incident responseHow to prepare
- Rehearse the first 90 minutes until it's automatic.
- Write three update templates: engineer, customer, executive.
- Drill common AI incidents: cross-tenant leak, stale index, cost spike, prompt injection, wrong citations.
- 8
Behavioural & ownership
45 min · stories
Ambiguity, ownership, conflict, escalations, delivery under pressure, leading without authority.
Senior signal: Creates structure in chaos, escalates early, and quantifies the outcome, in the first person.
Skill guide: Ownership & business valueHow to prepare
- Prepare 6–8 stories covering ambiguity, an incident, customer pain, production delivery and conflict.
- Each story needs a number and a lesson.
What interviewers score
Most rubrics score each dimension 1–5 and count only what you actually showed. The first dimensions carry the most weight. Senior answers aren't longer; they show better judgment about what to build, who owns it and what not to do.
- 1Problem discovery & clarification
- 2Business outcome & baseline
- 3Requirements & prioritisation (MVP, non-goals)
- 4Capacity & assumptions
- 5Architecture & integration (simplest reliable path)
- 6Data quality & retrieval
- 7Action safety (reversible vs irreversible)
- 8Security, permissions & compliance
- 9Reliability & observability
- 10Evaluation & quality measurement
- 11Rollout strategy
- 12Trade-offs & communication
- 13Ownership under ambiguity
Automatic fails
Any one of these can sink an otherwise strong round.
- No business outcome or success metric
- Security left to the prompt
- No evaluation plan
- Agent tools with no validation or approval
- No rollback plan, or a big-bang launch
- Promising 100% automation
- Can't explain your own code
Use it in every scenario round
The 8-step deployment framework
Step 1
Clarify context
Who's the customer, what triggered this, and what does success mean to them?
Step 2
Map the current system
Users, tools, data, and the workflow as it really runs today.
Step 3
Define success
A baseline, a target and a trust threshold, written as acceptance tests.
Step 4
Choose the integration path
Batch, webhook, events or embedded, chosen by how fresh data must be.
Step 5
Plan for failure
Timeouts, retries, idempotency, fallbacks and what the user sees when it breaks.
Step 6
Validate security & compliance
Identity, permissions outside the model, data handling and audit.
Step 7
Roll out gradually
Pilot cohort, gates, stop conditions and a rollback anyone can trigger.
Step 8
Communicate trade-offs
This choice, this benefit, this trade-off, this validation.
"It depends" is never an answer on its own. Say what it depends on, and what you'd do in each case.
Practise out loud
10 questions to rehearse
Discovery
A COO asks you to "put AI on the order-exceptions queue." What are your first five questions, and which answer would change your architecture most?
Integration
The customer's ERP only offers a nightly CSV and a SOAP API limited to 10 calls a minute. How do you build a near-real-time order-status feature?
RAG
Offline retrieval scores look fine, but users say answers miss the latest version of documents. What do you check, in what order?
Agents
When would you replace an LLM planner with a fixed state machine, even though the planner demos better?
Evals
Your LLM judge and your human raters disagree on 30% of cases. Which do you trust, and what do you change?
Security
How would you prove to a CISO that user A can never see user B's documents through your assistant, including through caches and logs?
Cost
Cut cost per resolved ticket by 60% this quarter without hurting quality. Walk through your plan and the metric that keeps you honest.
Incident
At 2 AM an agent starts sending duplicate emails to a customer's clients. What do you do in the first ten minutes?
Exec comms
In two sentences, explain to a CFO why the pilot will be draft-only for the first six weeks.
Behavioural
Tell me about a time you talked a customer out of the solution they asked for. What evidence did you use?
Record yourself, score each answer against the rubric above, fix your two weakest dimensions, then repeat the same question before moving on.