A strong 45–60 minute AI systems interview follows a repeatable arc: clarify → estimate → architecture → deep dives → production wrap. Memorize the skeleton so stress does not erase structure.

1. Clarify: users, modalities, latency SLO, accuracy bar, data sources, privacy, online vs batch (5–8 min)
2. Estimate: QPS, tokens in/out, cost/day, cache hit, peak load (5 min)
3. High-level architecture: client → API → retrieval/agent → model → tools/DB (10–12 min)
4. Deep dives: retrieval quality, tool safety, evals, failure modes (15–20 min)
5. Production: observability, guardrails, rollback, 10× scale (5 min)

Requirements → Numbers → Boxes & arrows
                              ↓
                    Pick 2 deep dives
                              ↓
                    Prod / failure / scale

FDE cases blend product sense, integration realism, and stakeholder craft. Lead with customer framing, then MVP, spike plan, and rollout — technology serves that narrative.

1. Frame the customer problem and constraints
2. Map stakeholders and success metrics
3. Scope an MVP with a Won't list
4. Spike the hardest integration risk first
5. Plan pilot rollout, training, and handoff

Problem → Stakeholders → MVP
              ↓
     Integration spike
              ↓
   Pilot → Feedback → Handoff

Senior interviews grade tradeoff quality. State options, criteria, recommendation, and what would change your mind. Avoid false dichotomies and "it depends" without criteria.

• Name 2–3 real options, not a strawman
• Pick explicitly: "I'd choose X because Y"
• List risks of your choice and mitigations
• Invite interviewer constraints that could flip the decision

Most AI lab/FAANG design questions collapse to a RAG-shaped system. Practice drawing a clean default diagram you can extend under pressure.

Client → API Gateway → Orchestrator
                           │
              ┌────────────┼────────────┐
              ▼            ▼            ▼
         Query rewrite  Retriever    LLM
              │         (hybrid)      │
              │            │          │
              │         Rerank        │
              │            │          │
              └────────────┴────► Prompt
                                   │
                              Answer+cites
Ingest: Connectors → Parse → Chunk → Embed → Index
         (ACL metadata on every chunk)

• Always draw ingest separately from query
• Label ACL filter on retrieve
• Mark eval harness and offline index quality jobs
• Add cache sitting beside the orchestrator when cost comes up

When the problem is open-ended (ops automation, research assistant, coding agent), draw a bounded plan–act loop with tool policy — not an unbounded brain.

User → Gateway → Agent runtime
                      │
                 Policy engine
                      │
            ┌─────────┼─────────┐
            ▼         ▼         ▼
         Planner   Tools API  Memory
            │         │         │
            └──── observe ──────┘
                      │
              step budget / stop
                      │
                 Final answer
Tools: search | ticket | code_exec (sandboxed)
Writes: require confirmation

• Write max steps and timeout on the board
• Separate read tools from write tools visually
• Show audit log as a first-class box
• Include human-in-the-loop for irreversible actions

Back-of-envelope estimates and crisp communication separate staff-level AI candidates. Practice token/cost math and saying uncertainty ranges out loud.

• State assumptions: "Assume 5 QPS peak, 2k in, 400 out…"
• Round aggressively; precision theater wastes time
• Compare options numerically when debating model tiers
• Summarize every 10 minutes: "So far we've decided X; next I'll dig into Y"