Interview Questions — Projects, Resume & Behavioral (GenAI)

Rounds: Resume/Projects (required) · Behavioral (required) Provenance: sections 1 and 3 are Generated practice; sections 2 and 4 are Common industry pattern (the kinds of probes GenAI project and judgment rounds often use). None is verified company-specific. Stories and project answers must be true — use only real work, real numbers and real ownership. Nothing below is pre-filled with achievements; fill the templates yourself.

1. Project-defense sequence (for each GenAI project on your resume)

Provenance: Generated practice. Walk the ladder; stop where you can’t answer without help — that’s the gap to fix.

1. Problem & users — what was broken, who cared, how was success defined?
2. Your ownership — which parts did you design/build vs the team?
3. Architecture — draw it; why each component?
4. Key decision + alternative — what did you reject and why (model, vector DB, chunking, framework)?
5. Evaluation — how did you know it worked? Dataset size, metrics, baseline, retrieval vs generation split?
6. Failure & debugging — worst production incident / hallucination case; root cause; fix.
7. Cost, latency, scale — real numbers, what you optimised.
8. Safety/security — injection, PII, permissions.
9. What you’d change — with 3 more months / 10× traffic.
10. Impact — measurable outcome (only if you can source it).

2. Probing questions about GenAI projects

Provenance: Common industry pattern. Each probe is worth answering with a real number or a real decision. Related lessons: Observability and eval pipelines, Evaluating LLM systems.

• How did you evaluate beyond “it looked good”? What was your baseline?
• What was your retrieval recall before/after your change?
• Why that chunk size / embedding model / vector DB?
• What did the system do when it didn’t know?
• How did you handle prompt injection / PII / permissions?
• What was your p95 latency and cost per query? What moved it most?
• When did the LLM fail in production, and how did you find out?
• What would you not use an LLM for in this system?

3. Behavioral story bank (fill with real experiences)

Provenance: Generated practice.

Rules: one story can serve multiple themes; keep to ~2 minutes; state your actions (“I”), quantify results only if true; end with what you learned. Practise each story out loud, then have someone probe it with follow-up questions.

4. GenAI-specific judgment questions (behavioral-technical hybrids)

Provenance: Common industry pattern. Related: Fine-tuning vs RAG vs prompting, Observability and eval pipelines.

• “A stakeholder wants to launch an AI feature next week with no eval set. What do you do?”
• “How do you explain to a non-technical PM why the model hallucinates?”
• “How do you decide a use case is not a good fit for an LLM?”
• “Describe how you’d roll out an LLM feature safely (shadow → canary → full).”