AI Engineering — Interview Notes

Concept write-ups for AI-engineering interviews, worked from the question list in amitshekhariitbhu/ai-engineering-interview-questions. The questions are theirs; the answers here are mine, written out rather than linked.

These cover the concepts that get asked and that I couldn’t already answer cold. Where a concept maps to something I actually built — Bubble or the Jira coding agent — there’s a callout saying what I did and why, because that’s the version of the answer that holds up when an interviewer probes.


The notes

NoteCovers
LLM InternalsQuantization · KV cache, prefill vs decode, why the first token is slow · why context windows are limited · temperature, top-k, top-p · positional encoding and RoPE · Mixture of Experts
Fine-Tuning and AlignmentFine-tune vs RAG vs prompt · full fine-tuning vs LoRA vs QLoRA · RLHF, PPO, reward hacking, alignment tax · DPO and GRPO
RAG AdvancedNaive vs agentic vs Self-RAG · HyDE, decomposition, step-back · multi-hop · GraphRAG · per-user access control · parent-child chunking
AI AgentsMCP · ReAct vs Plan-and-Execute · how tool calling actually works · agent memory and eviction · prompt injection and agent security · loop termination
Embeddings and Vector SearchDimensionality and cost · embedding quantization and Matryoshka · sparse vs dense and hybrid search · changing the embedding model in production · metadata and multi-tenancy
EvaluationLLM-as-judge and its four biases · deterministic checks · offline vs online · golden datasets · red teaming · reading benchmarks

Suggested reading order

  1. LLM Internals — everything else assumes it. KV cache and prefill/decode in particular explain latency questions across the rest.
  2. RAG Advanced and Embeddings and Vector Search — read as a pair; they’re two halves of the same system.
  3. AI Agents — the section with the most of my own project evidence.
  4. Evaluation — the identified weak spot, so worth more than one pass.
  5. Fine-Tuning and Alignment — least likely to come up for an applied engineering role, but “fine-tune or RAG?” is asked constantly.

Already covered elsewhere

Not repeated here — these live in the round-2 prep material: self-attention, HNSW vs IVFPQ, distillation, chunking strategies, bi-encoder vs cross-encoder reranking, faithfulness and context recall, lost-in-the-middle, RAG debugging, document freshness, human-in-the-loop, precision/recall and threshold selection, MLOps, voice agent architecture.

Adjacent notes from actual deployments: AWS · Docker


How to use these

Reading these is not the same as being able to answer. The failure mode is real and I’ve hit it — reading an explanation eight hours before an interview and still drawing a blank on the question.

So: for each concept, try to say the answer out loud first, then read. The gap between what you produced and what’s written is the only part worth studying.