AI System Design Interview Questions
Dashboard
0%
1
Curious builder0 XP earned · 300 to level 2
0 daysFinish a lesson to begin
Badge collection0 of 6 unlocked
51 small wins to finish your path

Q50HardScenario

Mock round: "Design an AI assistant for a bank's retail customers." Show how you'd run the 45 minutes.

30-second answerSay your answer out loud first, then reveal.

Minutes 0–5: Clarify

  • Channels: mobile app chat, WhatsApp, voice? Languages: English, Hindi, regional?
  • Scope v1: FAQs + account information (balance, transactions) + service requests (card block, statement)? Transfers in v1? (Suggest no, or only with strict confirmation flows.)
  • Constraints: RBI/regulatory guidelines, data residency, audit, accessibility.
  • Success: containment rate, CSAT, call-centre deflection, zero unauthorised actions.

Minutes 5–8: Estimates. 20M customers, 2M monthly active users of chat, ~10 sessions/user/month → ~670K sessions/day. Peak at salary days. Token and cost estimates; most traffic is simple intents, which argues for a router.

Minutes 8–18: High-level design

A customer's message passes an authenticated session and input guardrails to an intent router that sends account questions to a read-tool assistant, FAQs to RAG over approved docs, complaints to a human, and card actions to a deterministic workflow with step-up auth; replies pass output guardrails and every turn is audit logged.

Minutes 18–33: Deep dives (examples)

  • Authorisation: tools receive the customer ID from the session, never from the model; read-only scopes; transactions only via deterministic flows with OTP/biometric step-up.
  • Hallucination control: answers about rates and fees come only from retrieved approved documents, with citations; if not found, say so and escalate; no financial advice beyond approved scripts.
  • Prompt injection: transaction descriptions and uploaded documents are untrusted, and no tool can move money based on model output alone.
  • Compliance: audit trails, retention, data residency in India, consent, complaint escalation rules.

Minutes 33–45: Evals, monitoring, rollout, trade-offs

  • Golden sets per intent, red-team suites (social engineering, injection), multilingual evals.
  • Shadow → internal staff → 1% customers → gradual expansion; kill switches per intent.
  • Trade-offs discussed: coverage vs risk (start narrow), latency vs guardrail depth, API vs self-hosted given data constraints.
What this demonstrates. Structure, safety-first judgement in a regulated domain, and clear scoping. That's what interviewers grade most heavily.
PreviousDesign an observability and tracing platform for LLM applications handling billions of spans per day.

Every expert started right here.