AIシステム設計面接キット

Reference latency and unit tables, assumptions register, QPS, storage, bandwidth, cache, server and cost formulas, worked example, sensitivity, traceability

Problem framing, metrics, data, features, baselines, training, offline and A/B evaluation, serving, monitoring, rollout

Recommender funnel: two-tower retrieval and ANN, pre-ranking and ranking, feature freshness, training, exploration, A/B tests, feedback loops and bias

Real-time payment fraud scoring: latency budget, streaming features, rules plus ML, review queues, label delay, selection bias, model monitoring

Enterprise RAG search design: connectors, document and chunk ACL filtering, permission sync, hybrid retrieval, RRF, reranking, citation verification, tenancy

Chat assistant app layer: resumable SSE streaming, conversation store, context budget and summaries, uploads, tools, moderation, rate limits, residency

AI coding assistant design: IDE and CLI clients, repo indexing, context assembly, agent loop, microVM sandbox, MCP and Agent Skills, approvals, security

Customer support AI agent design: channels, intent routing, policy retrieval, refunds with limits and approvals, human handoff, metrics, safety, audit

Real-time voice agent design: SIP and WebRTC ingress, streaming ASR, turn detection, barge-in, LLM with tools, streaming TTS, latency budget, consent, failover

LLM gateway and router design: unified API, provider adapters, token quotas, routing and fallback, caching, PII redaction, guardrail hooks, metering

Gateway, router, continuous batching, paged KV cache, GPU memory and throughput math, TTFT/TPOT SLOs, autoscaling, quantization, HNSW/IVF, cost per token

Agent run platform design: durable step journal, resume, MCP tool broker with Tasks, A2A calls, human approvals, tenant budgets, evaluation hooks

Code execution sandbox for agents: microVM vs gVisor vs containers, egress proxy, credential broker, quotas, snapshot restore, warm pools, abuse detection