AI Interview Questions

Advanced scenario question on why LLMs hallucinate — next-token prediction vs fact verification, root causes, and production defenses asked at OpenAI, Google, Meta, Anthropic, Microsoft, and Amazon.

Part 2 of the LLM hallucination handbook — factual, citation, reasoning, math, code, temporal, context, and tool-use hallucinations; decoding risks; RAG failure modes; production case study.

Part 3 of the hallucination handbook — production defense layers from retrieval and prompting through guardrails, citations, confidence scoring, eval frameworks, and enterprise architecture. Asked at OpenAI, Google, Meta, Anthropic, Microsoft, and Amazon.

Scenario interview on defining and communicating hallucination metrics to executives.

Hard production LLM question on routing, cascading, and router evaluation.

Model-based interview on DPO vs RLHF — preference pairs, training simplicity, and production trade-offs.

Scenario interview on safe GenAI rollouts — flags, shadow eval, canaries, and kill switches.

LLM interview topic on distillation — teacher/student training, latency/cost wins, and quality regression risks.

Hard project question on LLM gateway — auth, routing, rate limits, logging, key management, and multi-provider abstraction.

Scenario question on cross-model tokenizer bugs — chunk boundaries, context overflow, and embedder vs LLM alignment.

Model question on log probabilities — uncertainty signals, calibration limits, and confidence workflows.

Hard project question on LLM red-teaming — adversarial prompts, automation, severity rubrics, and fix-verify cycles.

Scenario question on offline golden eval vs online A/B, feedback, and guardrail metrics in production.

Project question on semantic cache — embedding similarity hits, TTL, invalidation, and staleness risks.

Project question on LLM routers — intent classification, cascades, early exit, and quality monitoring.

Hard Amazon company question — model cascades, unit economics, SageMaker/Bedrock routing, and SLO-aware cost optimization.

Hard Microsoft company question — enterprise copilot grounding, Graph permissions, tenant isolation, and compliance.

Hard Anthropic company question — least-privilege tools, confirmation flows, injection defense, and constitutional guardrails.

Hard Meta company question — Llama serving, GPU clustering, quantization, batching, and open-weight ops.

Hard Google company question — document QA with OCR, layout, retrieval, multimodal models, and table/chart grounding.

Hard OpenAI company question — chat product system design with safety stack, eval gates, streaming, and grounding.

Hard scenario on agent memory — episodic vs semantic memory, vector stores, summarization, and model parametric limits.

Scenario question comparing CoT prompting with calculators, code interpreters, and search tools for reliable reasoning.

Project question on synthetic data — distillation datasets, diversity, contamination, quality filters, and human review.

Scenario question on abstention — retrieval scores, calibration, UX of uncertainty, and avoiding confident wrong answers.

Hard scenario on jailbreaks — DAN, prompt injection, alignment limits, red-teaming, and layered defenses.

Hard project question on use-case benchmarking — golden sets, task metrics, regression gates, and leaderboard limitations.

Hard model question on MoE — routed FFN experts, sparse compute, capacity factors, and Mixtral/GPT-4 class rumors.

Model question on speculative decoding — draft model, target verification, lossless speedup, and production deployment notes.

Scenario question on resilient LLM clients — exponential backoff, 429 handling, idempotency keys, and duplicate-safe writes.

Easy project question on SSE/token streaming — perceived latency, client backpressure, partial JSON, and disconnect recovery.

Project question on prefix/prompt caching and KV cache reuse — TTFT optimization and provider billing implications.

Hard scenario on PII in LLM systems — detection, redaction, DPA compliance, log minimization, and regional deployment.

Scenario question on LLM guardrails — prompt injection defense, output validation, NeMo/Guardrails patterns, and policy layers.

Hard project question on INT8/INT4 quantization, knowledge distillation, and serving smaller models for cost and latency.

Company/scenario question on open vs closed LLMs — self-hosting economics, compliance, upgrade velocity, and hybrid strategies.

Hard model question comparing CAI, RLHF (PPO), and DPO — data efficiency, stability, and production fine-tuning trade-offs.

Model question on RLHF — SFT, reward model, PPO alignment, and how human preferences change helpfulness vs hallucination trade-offs.

Model question on multimodal LLMs — image/audio tokenization, document QA, OCR pipelines, and latency/cost trade-offs.

Project question on JSON schema mode, tool/function calling, validation loops, and agent reliability patterns.

Hard scenario on LLM cost control — token budgeting, routing, caching, batch APIs, and unit economics at millions of queries.

Scenario question on LLM latency — prefill vs decode, retrieval overhead, streaming, caching, and model routing fixes.

Hard project question on LLM-as-judge — rubric design, position bias, human calibration, and production eval pipelines.

Scenario question on classical NLP metrics — what they measure, where they break on paraphrase and factuality, and what to use instead.

Model question on dense vector embeddings — semantic similarity, clustering, classification, dedup, and recommendation beyond vector search RAG.

Decoding parameters interview question — how sampling controls randomness, creativity, and hallucination risk in production.

Scenario question on context windows — token budgets, lost-in-the-middle bias, and why RAG still matters despite 1M-token claims.

Conceptual Transformer interview answer — self-attention, encoder-decoder vs decoder-only, and why parallel training beat RNNs for language modeling.

Foundational model question on BPE/tokenizers — why token counts differ from words, and production impact on cost, limits, and RAG chunking.

Scenario question comparing prompting, RAG, and fine-tuning for domain knowledge — freshness, cost, auditability, and when to combine approaches.