0/32
Not saving
01
Interview prompt
Design a GenAI FDE (Forward Deployed Engineer) solution for Executive Dashboard Assistant. The customer has ambiguous requirements, sensitive data, and expects a production rollout, not a demo.
02
Clarify the customer problem
- What business process is broken today?
- Who are the users and what decisions do they make?
- What is the cost of delay, error, or manual work?
- What data is available, trusted, missing, or restricted?
- What does success mean in the first 30, 60, and 90 days?
03
Users and workflows
| User | Workflow | Current pain | AI assist opportunity | Human approval needed? |
|---|---|---|---|---|
| Primary operator | ||||
| Manager/reviewer | ||||
| Admin/security |
04
Requirements
Functional
Non-functional
- Latency target
- Availability target
- Cost budget
- Security/privacy constraints
- Audit/compliance requirement
05
Data and integration map
| Data source | Format | Owner | Freshness | Permission model | Risk |
|---|---|---|---|---|---|
06
Proposed architecture
Fill in each part of your architecture for this case:
- Ingestion
- Retrieval
- Agent/tool use
- Human approval
- Evaluation
- Monitoring
07
Evaluation plan
| Metric | Good threshold | Bad threshold | Test dataset | Owner |
|---|---|---|---|---|
| Answer groundedness | >= 90% | < 80% | Golden Q&A | FDE/SME |
| Permission violations | 0 | > 0 | Red-team set | Security |
| Citation accuracy | >= 95% | < 85% | Sample docs | SME |
| Task completion | >= 80% | < 65% | Workflow tests | Product |
08
Failure modes
- Bad retrieval or missing context
- Hallucinated policy or procedure
- Cross-tenant or role-based data leakage
- Unsafe tool call or write-back action
- Stale documents or bad sync
- Cost/latency spike
09
Rollout plan
- Offline prototype using historical cases.
- SME review and evaluation set creation.
- Read-only pilot with citations.
- Human-approved actions.
- Limited production canary.
- Monitor, tune, and expand.
10
Weak vs strong answer
Weak: “I would use RAG (retrieval-augmented generation) with a vector database and ask GPT to answer.”
Strong: “I would first define the operational workflow and risk boundary, then design permission-aware retrieval with grounded citations, evals, red-team tests, human approval for risky actions, and production observability before rollout.”
11
Candidate scorecard
The closer
Final 2-minute spoken answer
Rehearse this out loud before you look at it. Summarising the whole design in two minutes is its own skill, and it is usually the last thing you are asked for.