Roadmaps / AI / Interview topics
Reference
Interview topics
Use these for AI engineering and backend roles. Each follow-up is the question a good interviewer asks next. The Staff line is what lifts the answer.
Phase 1
| Question | Follow-up | A Staff answer adds |
|---|---|---|
| How do you cut the cost of an AI feature? | What do you give up? | A measured table of cost against quality, caching rules and a router |
| Where does latency come from? | How do you reduce it? | Time to first token against generation, streaming and shorter inputs |
| How do you handle a model timeout? | What about retries? | Backoff with jitter, idempotency and a cap on total spend |
| Hosted or self-served model? | How do you decide? | Volume, latency, data rules and a break-even calculation |
System design prompt (20 minutes): a summarisation feature for 1 million documents a day. Outline: numbers and budget (3 min), model choice and batching (5), failure handling (4), quality checks (4), observability (4).
Phase 2
| Question | Follow-up | A Staff answer adds |
|---|---|---|
| How do you know the system got better? | What if the score changes by 2 points? | Spread over runs, a noise floor and a paired comparison |
| How do you build an evaluation set? | How do you keep it honest? | Coverage, unanswerable cases, a second labeller and a held-out part |
| Can a model judge another model? | When does it fail? | Agreement with human labels, bias and a re-check after every change |
| How do you stop regressions? | What runs in CI? | A fast subset in CI, a full run nightly and a rule for failing a build |
System design prompt (20 minutes): an evaluation pipeline for a customer-facing assistant. Outline: what to measure (3 min), data and labels (5), scorers (5), CI and monitoring (4), review process (3).
Phase 3
| Question | Follow-up | A Staff answer adds |
|---|---|---|
| How does retrieval-augmented generation fail? | How do you find the cause? | A failure list by cause, recall at k and a fix with a measurement |
| Vector, keyword or hybrid search? | When do you add a reranker? | A comparison on your questions and the latency cost |
| How do you cite sources? | How do you check a citation? | Passage ids, a checker and a refusal rule |
| How do you handle access control? | Where do you filter? | In the retrieval query, with a test for a second user |
System design prompt (20 minutes): a company knowledge assistant with per-team access. Outline: requirements and numbers (3 min), ingestion and chunking (5), retrieval and permissions (5), answer generation and citations (4), freshness and cost (3).
Phase 4
| Question | Follow-up | A Staff answer adds |
|---|---|---|
| When do you use an agent? | When do you not? | A workflow by default, step limits and a measured reason for the loop |
| How do you defend against prompt injection? | Can a filter solve it? | Least privilege, confirmations, an allow-list and a suite of attacks |
| How do you design a tool server? | What about authorisation? | Narrow tools, validated arguments, per-request identity and audit logs |
| How do you upgrade a model? | What can break? | An evaluation comparison, an injection re-run and a staged rollout |
System design prompt (20 minutes): an agent that can refund small orders. Outline: limits and risk (3 min), tools and confirmations (5), injection defences (5), audit and monitoring (4), rollout (3).