GenAI Engineering Interview Prep
Test & Track Your Readiness

What the modern AI engineer stack looks like and what interviewers expect you to know. The full mental map: LLMs, RAG, agents, evals, fine-tuning, safety, and production ops. Where each domain sits and why it matters. The competency model companies hire against.

Transformer architecture from first principles — attention heads, positional encoding, residual streams, MLP layers. How pretraining, SFT, and RLHF shape model behaviour. Inference: batching, KV cache, speculative decoding, quantisation. What every AI engineer must be able to explain in an interview without hand-waving.

Beyond basic prompting: chain-of-thought, self-consistency, step-back, multi-turn discipline, system prompt design for reliability, few-shot exemplar construction, prompt versioning, and regression testing. The interview-critical prompting patterns that separate junior from senior AI engineers.

What goes into the context window, in what order, and why. Context rot, lost-in-the-middle at scale, dynamic context construction strategies, RAG vs long-context economics, compression techniques (summarisation, extraction, distillation), context budgets, and the interview questions companies use to probe this skill.

How to design a complete RAG pipeline end-to-end in a 45-minute system design round. Indexing pipeline, embedding strategy, vector store selection, retrieval, reranking, prompt assembly, generation, and evaluation hooks. Common follow-up questions and the trade-off answers that impress interviewers.

Deep system design for retrieval at scale: HNSW vs IVF index structures, the recall-latency-cost triangle, embedding model selection criteria, multi-vector representations, late interaction (ColBERT), namespace strategies, filtering at scale, embedding versioning and migration. Design questions with worked answers.

Beyond basic dense retrieval: hybrid search with RRF fusion, cross-encoder reranking, HyDE (hypothetical document embeddings), query decomposition, step-back prompting for retrieval, GraphRAG and knowledge graphs, structured-data RAG (Text-to-SQL), multi-hop retrieval. When each wins, interview-grade trade-off analysis.

Building an evaluation system that doesn't lie: golden dataset construction, LLM-as-judge patterns and biases, RAGAS metrics (faithfulness, answer relevance, context precision/recall), human-eval calibration, regression detection in CI/CD, A/B testing for LLM apps, and the evaluation questions interviewers ask.

The canonical patterns every AI system design interview expects: Router-Dispatcher, Orchestrator-Subagent, RAG-Augmented Pipeline, Guardrail Sandwich, Async Evaluation Loop, Cache-then-Generate, Model Gateway. Each pattern with when to use, scaling behaviour, failure modes, and interview whiteboard diagrams.

Designing GenAI systems for 100K+ concurrent users: request routing, tenant isolation, rate limiting strategies, model load balancing, autoscaling serving infrastructure, fallback chains, burst handling, multi-region LLM serving, the 10× cheaper request path. Real architectures from companies at scale.

The decision matrix every AI engineer must internalise: when to fine-tune, when to RAG, when to just prompt better. Dataset requirements, cost models, latency impacts, quality ceiling analysis, distillation from frontier models, PEFT vs full fine-tune, and the interview trade-off questions around each choice.

How alignment actually works at a technical level: reward modelling, PPO for RLHF, DPO as a simpler alternative, Constitutional AI (CAI), rejection sampling, and RLAIF. What alignment buys and what it doesn't. The interview questions around bias, alignment tax, and what 'aligned' actually means in production.

Designing systems that process text, images, audio, and video together: vision-language models, OCR and document understanding, audio transcription pipelines (Whisper), video understanding at scale, multimodal RAG, adversarial image injection risks, latency and cost profiles per modality, and multimodal system design interview patterns.

Production-grade guardrail architecture: input classifiers, output validators, content moderation pipelines, toxicity detection, PII detection and redaction, refusal handling, the guardrail sandwich pattern, latency cost of safety layers, guard model choices (Llama Guard, Azure AI Content Safety, Perspective API), and the responsible AI interview lens.

OWASP LLM01 in full depth for an interview audience: direct vs indirect injection, RAG corpus poisoning, multi-modal injection via images, tool-result injection in agent loops, trust tiering for untrusted content, defense-in-depth strategies, red-team suites (Garak, PyRIT), and the architectural patterns that survive adversarial inputs.

Running LLM applications in production without flying blind: OpenTelemetry semantic conventions for LLMs, distributed tracing across multi-step pipelines, cost attribution, quality alerting, latency SLOs for LLM apps, model failover automation, prompt versioning in CI/CD, evaluation gates, and the observability tooling landscape (LangSmith, Phoenix Arize, Braintrust, Helicone).

The economics of LLM applications in production: input vs output token cost asymmetry, prompt caching ROI, semantic caching design, model-tier routing, batching strategies, speculative decoding for latency, streaming first-token-time optimisation, the sub-$0.01/request target, and interview questions on cost engineering.

Full-stack architecture for production GenAI products: the edge-to-model request path, streaming SSE architecture, client-side state management for streamed responses, async job pipelines for long-running AI tasks, feature flagging for model rollouts, the multi-model product architecture, and real case studies from companies like Notion, Linear, Cursor, and Perplexity.

The questions interviewers ask about reasoning models (o-series, Claude extended thinking, DeepSeek R1) and agentic systems: when to use reasoning models vs standard models, budgeting thinking tokens, tool use with reasoning, the agent loop design, multi-agent coordination, and the system design interview for an agentic customer support system.

The consolidated interview playbook: the 30 most common GenAI technical questions with model answers, the system design framework for AI-specific rounds (different from classic HLD), how to structure trade-off answers, the vocabulary and signal words senior interviewers look for, and the most common mistakes candidates make in AI engineer interviews.