AI Prep Buddy — Day-of-Interview Cheat Sheet

The condensed version. Read this in the 20 minutes before your interview, not the full 1,801-question bank. Organized around the questions you’re most likely to actually get.

The 5 Anthropic Agent Patterns (memorize the names)

Prompt Chaining → Routing → Parallelization → Orchestrator-Worker → Evaluator/Optimizer. If asked “how would you architect an agent,” name the pattern that fits before describing it. Default answer to “should I build an agent for this”: “find the simplest solution possible — often this means not building agentic systems at all.”

System Design: The Universal Opening Move

For any system-design question: clarify scale → state your latency/cost budget → sketch data flow → call out the failure mode you’re most worried about → name the eval/monitoring plan. Interviewers grade the process, not whether you land on the “right” architecture.

The 6 Patterns That Answer 80% of System-Design Questions

1. Two-stage retrieve-then-rank (RAG, search, recsys) — cheap broad candidate generation → expensive precise re-ranking on a shortlist.
2. Gateway + fallback chain — one abstraction layer, provider-agnostic, automatic failover, never single-source a dependency.
3. Eval-gated CI/CD — golden dataset → shadow → canary → full rollout, with rollback at every stage.
4. Human-in-the-loop risk tiering — auto-execute low-risk/reversible actions, gate high-risk/irreversible ones.
5. Online + offline feature store from one definition — eliminates training/serving skew by construction.
6. Defense-in-depth guardrails — input filter → hardened system prompt → output filter → structural validation. No single layer is ever sufficient alone.

Post-Training, Current State (say this, not just RLHF)

SFT → alignment stage. For reasoning capability: GRPO + RLVR is now the frontier default (DeepSeek-R1, GPT-5.3, Nemotron 3) — verifiable rewards (math/code correctness), no separate reward model, no critic network. For general preference alignment (tone, safety, helpfulness): DPO/RLHF still used. Don’t just say “RLHF” as the whole answer in 2026 — it reads as dated.

Inference/Serving, Current State

vLLM = production default (PagedAttention, broad compatibility). SGLang = real competitor now, wins on prefix-heavy workloads (chat/RAG/agents) via RadixAttention. TensorRT-LLM = max throughput, single stable model, NVIDIA lock-in, compile step. TGI = deprecated (Dec 2025). If asked “what would you deploy today” — mention SGLang, not just vLLM.

The One-Liner Definitions You’ll Get Asked Cold

• Why scale attention by √d_k: prevents large dot products from pushing softmax into vanishing-gradient regions as dimension grows.
• Why GQA over MHA: cuts KV-cache memory/bandwidth (the real serving bottleneck) while keeping most of MHA’s quality.
• Why KL penalty in RLHF/DPO: stops the policy from drifting so far from the reference model that it reward-hacks.
• Why LoRA works: the update to pretrained weights for a new task is usually low-rank, so you don’t need full-rank trainable weights.
• Bias-variance tradeoff: high bias = underfits (too simple); high variance = overfits (too sensitive to training noise).
• Why RAG can worsen hallucination: bad retrieval → confident answer grounded in the wrong context. RAG only helps if retrieval quality is genuinely high.

Real Incidents Worth Citing (stronger than hypotheticals)

• Air Canada (2024): chatbot invented a refund policy; tribunal held the airline legally liable. → Companies own their AI’s claims.
• Cyera Research (May 2026): 344 verified enterprise-harm cases from AI agents, 188 with zero external attacker — the agent did it alone. → Excessive agency is a live risk, not theoretical.
• McDonald’s/Paradox.ai breach (2025): 64,000 applicants exposed via a weak password + IDOR, not a sophisticated AI attack. → AI vendor due diligence = standard security hygiene too.

Governance Framework Names (drop these, don’t over-explain)

NIST AI RMF (Govern/Map/Measure/Manage) · ISO/IEC 42001 (AI management system, not a one-time checklist) · EU AI Act high-risk obligations (currently a live compliance deadline — Annex III enforcement was slated for Aug 2, 2026, with a proposed-but-not-fully-enacted delay to Dec 2027 in play; say “the exact date is in flux, but the obligation categories are what matter architecturally” if pressed) · SR 11-7 (banking model risk — independent validation + ongoing monitoring).

Behavioral: The Structure That Actually Works

STAR, but the Action should show judgment under ambiguity and the Result should include what you’d do differently — interviewers weight self-awareness about failure more than a clean success story.

Questions to Ask Them (don’t skip this)

One specific, current question about their architecture/research direction — not about perks or generic “what’s the culture like.” Shows you did homework, not just interview prep.

If You Freeze

Say the pattern name out loud even if you haven’t fully worked out the details — “this looks like a two-stage retrieve-then-rank problem, let me think through the ranking stage” buys you thinking time and shows structured reasoning, which is what’s actually being graded.

Frontier Topic Quick-Reference (Sections 32–47)

• Multimodal: early fusion (one model, all modalities) vs late fusion (separate encoders → projection). CLIP aligns image-text via contrastive loss. For multimodal RAG, you need layout-aware parsing + table extraction, not just text chunking.
• Fine-tuning decision: prompt engineering first → RAG if you need knowledge → LoRA/QLoRA if you need behavior/style → full fine-tune only if you have massive domain-specific data and compute budget.
• Quantization quick hierarchy: FP16 → INT8 (PTQ, easy, ~1% quality loss) → INT4 (AWQ/GPTQ, noticeable on non-English) → speculative: FP4/FP8 (QAT required). GGUF for CPU/edge.
• Fairness impossibility: you cannot simultaneously satisfy demographic parity, equalized odds, and calibration — pick the one that matches your application’s harm model.
• Diffusion models: noise → denoise over T steps. Classifier-free guidance trades diversity for prompt adherence. DiT (Diffusion Transformer) is replacing U-Net as the backbone.
• Voice agent latency budget: ASR (~200ms) + LLM (~300ms) + TTS (~200ms) = ~700ms minimum. Streaming ASR + speculative TTS start before LLM completes to hit conversational feel.
• Edge AI: quantize → prune → distill → use hardware delegates (CoreML/NPU). Federated learning = model goes to data, not data to model.
• Distributed training: ZeRO-1 shards optimizer, ZeRO-2 adds gradients, ZeRO-3 adds parameters. Tensor parallelism splits layers, pipeline parallelism splits stages. Memory ≈ 18× params (FP32 Adam).
• Data-centric AI: “more data” is rarely the right answer — “better data” almost always is. Synthetic data works when verified by a stronger model or deterministic check.
• Search/Ranking: two-stage always — cheap recall (BM25 + embedding) → expensive precision (cross-encoder re-ranker). ColBERT = late interaction compromise between bi-encoder speed and cross-encoder quality.
• Causal inference: correlation ≠ causation. Use DAGs to identify confounders. Propensity score matching when you can’t randomize. Uplift modeling when you want to target the persuadable, not the already-converted.
• Agentic Systems: ReAct for short-horizon tool execution → Tree-of-Thought / MCTS for complex search space planning. Always sandbox execution in MicroVMs (gVisor/Firecracker) with strict eBPF syscall limits.
• Hardware Acceleration: Memory wall is the primary LLM decode bottleneck (HBM bandwidth bound). Use Triton for custom fused kernels. FlashAttention-3 avoids HBM writes via online softmax & SRAM tiling.
• AI Security & Red Teaming: Dual-LLM architecture separates untrusted data parsing from privileged reasoning. Defense-in-depth requires output guardrails (Llama Guard/NeMo) + Confidential Computing (GPU TEEs).
• Long-Context Mechanics: SSMs (Mamba) offer $O(N)$ linear inference state tracking vs Transformer $O(N^2)$ quadratic KV-cache growth. PagedAttention eliminates memory fragmentation; SGLang RadixAttention enables dynamic tree prefix reuse.
• Domain-Specific AI: Robotics uses Vision-Language-Action (VLA) models mapping visual inputs directly to continuous motor actions. Healthcare AI demands MedMedQA evals & HIPAA PHI de-identification. Financial AI relies on microsecond-level limit order book (LOB) models. Software agents combine AST Code Property Graphs with test-driven execution loops.
• Cloud AI Deployments: AWS uses Bedrock Provisioned Throughput + SageMaker Async + EKS Karpenter GPU autoscaling + Inferentia2; Azure uses Azure OpenAI PTU + Azure ML vLLM Online Endpoints + AKS KEDA (queue/latency scaling) + APIM Gateway; GCP uses Vertex AI Endpoints + Cloud Run GPU (serverless L4) + GKE TPU/GPU Ray clusters + AlloyDB pgvector; FinOps uses Prometheus DCGM GPU utilization exporters + Spot/Preemptible reserved capacity strategies.

Full depth on every topic here: README.md (questions) → answers.md (frameworks) → diagrams.md + patterns.md (diagrams) → sources.md (citations).