Interview prep
AI Engineer: 7-day sprint
Seven days on what AI and LLM engineer loops test - how LLMs work, adaptation, RAG, agents, evaluation, inference, and security.
Rapid-fire this sprint0/27 done
Reading mode for lessons:
The job. Build products on top of foundation models: prompts, retrieval, tools and agents, fine-tuning when it pays off, evaluation, and the serving, cost, and safety work that makes it production-grade.
What the loop usually tests
| Round | What they probe | Where it's covered |
|---|---|---|
| LLM fundamentals | Tokens, attention, decoding, context windows, why models hallucinate | Day 1 |
| Adaptation | Prompting vs RAG vs fine-tuning, LoRA, preference tuning | Day 2 |
| LLM system design | RAG, agents, evaluation, guardrails, cost and latency | Days 3 to 6 |
| Inference | KV cache, batching, quantization, latency metrics | Day 6 |
| Practical coding | Build a small RAG or tool-calling loop, write an eval | Each lesson's implementation challenge |
Day 1: How an LLM produces text
- Read: Tokenization and pretraining, Self-attention, Transformer architecture, Decoding
- Also read: Attention variants
- Labs: tokenizer, self-attention, temperature/top-k/top-p
- Cheat sheet: transformer architecture
- Be ready to: explain why token count drives cost and latency, what temperature does to the distribution, and what the KV cache stores.
Day 2: Adapting models
- Read: Adapting LLMs, Fine-tuning in practice
- Also read: Prompt engineering, Few-shot and chain-of-thought
- Lab: LoRA and fine-tuning memory
- Cheat sheets: LLM adaptation, fine-tuning
- Be ready to: argue prompting vs RAG vs fine-tuning for a concrete product, compute LoRA's trainable parameters, and explain DPO in two sentences.
Day 3: Retrieval
- Read: The RAG pipeline, Vector search, Text to vectors
- Also read: Context engineering
- Labs: RAG pipeline, approximate nearest neighbors, TF-IDF
- Cheat sheets: RAG pipeline, vector search
- Be ready to: debug "the right document exists but the answer is wrong", justify hybrid search, and pick chunking settings.
Day 4: Tools and agents
- Read: Tool use and agents, Reliable agents
- Also read: Agent memory, Model Context Protocol, Multi-agent systems
- Lab: agent loop
- Cheat sheet: agents
- Be ready to: explain how a tool call works end to end, when a fixed workflow beats an agent, and how per-step reliability compounds.
Day 5: Evaluation and security
- Read: Evaluating LLM systems, Securing AI systems
- Lab: prompt injection
- Cheat sheets: LLM evaluation, security controls
- Be ready to: design an eval set and an LLM-as-judge rubric, prove a prompt change helped, and stop indirect prompt injection in code rather than in the prompt.
Day 6: Inference, serving, and production
- Read: LLM inference, Serving LLMs, Production architecture, Reliability, cost, and observability
- Also read: Mixture of experts, Open-model architectures
- Labs: KV cache and inference memory, static vs continuous batching, production architecture
- Cheat sheets: LLM inference, production reliability
- Be ready to: estimate KV-cache memory, explain TTFT vs time per output token, and cut serving cost without losing quality.
Day 7: Mock interview
- Rapid-fire: 20 mixed cards from this sprint, 60 seconds each (open the drill).
- Design, 40 minutes each, out loud: Design a customer-support assistant that answers from 50,000 help articles and can issue refunds. Then: Our LLM feature costs USD 40k a month and p95 latency is 9 seconds. Cut both in half.
- Coding, 25 minutes: a minimal tool-calling loop with a JSON schema, a step limit, and error handling; then a 20-question eval harness that scores it.
- Project story: the RAG evaluation workbench or the agent sandbox, told in five minutes with numbers.