Builds production GenAI applications, prompting, RAG, agents, fine-tuning, inference cost and latency tradeoffs.
Learn how text becomes tokens, tokens become vectors, vectors get mixed by attention, and how a full transformer block fits together.
By the end of this week, you can explain scaled dot-product attention from scratch on a whiteboard and walk through a full transformer block in 10 minutes.
Every senior LLM interview opens here. If you can't explain attention from scratch, the rest of the loop won't matter.
Watch out: Most students conflate self-attention with cross-attention. If you can't explain the difference cleanly, you are not done with this week.
Learn how a user question becomes a retrieval, then an answer with citations: chunking, vector databases, hybrid retrieval, and evaluation.
By the end of this week, you can sketch a full RAG pipeline and defend your chunking strategy under interview pressure.
Most production LLM systems are RAG under the hood. Interviewers probe your trade-off reasoning between chunk sizes, recall, and cost.
Watch out: Choosing chunk size by gut feel is the most common wrong answer. Tie every chunking decision to a concrete recall, precision, or cost trade-off.
Learn when and how to fine-tune: LoRA, RLHF, DPO, data preparation, and evaluation.
By the end of this week, you can argue when NOT to fine-tune and walk through LoRA, RLHF, and DPO from memory.
Knowing when NOT to fine-tune is half the answer. Interviewers want to see you reach for prompting and RAG before reaching for compute.
Watch out: Candidates often reach for fine-tuning when a better prompt would have worked. Prove you know the cheaper alternatives first.
Build agents with tool use and MCP, then deploy with inference optimization, observability, and scaling.
By the end of this week, you can design an agent loop with retries, tool errors, and a real latency budget on a whiteboard.
Agents and serving are where most candidates wave hands. Concrete latency, retry, and tool-error stories make you stand out.
Watch out: Hand-wavy 'the agent decides' answers fail every senior loop. Bring concrete numbers and failure modes.
Design a complete LLM serving stack from scratch and prove how you would measure it offline and online.
By the end of this week, you can design a full LLM serving stack on a whiteboard and defend your evaluation strategy at both offline and online levels.
System design rounds tie everything together. Reasoning about evaluation (offline + online) separates senior candidates from mid-level.
Watch out: Most candidates skip evaluation and pay for it. Always lead with how you would measure the system's wins and losses.