Month 12-Week 2: Interview prep-coding, ML system design, breadth, behavioral¶

Week summary¶

Goal: Targeted prep for 4 interview formats: coding, ML system design, ML breadth, behavioral. Don't over-prep-but don't show up cold.
Time: ~9 h over 3 sessions.
Output: Format-specific notes; 5 STAR stories rehearsed; 1 mock interview.

Why this week matters¶

A year of building means little if you can't articulate it under pressure. One focused week of prep is high-leverage and prevents the gap between "I built it" and "I can talk about it."

Prerequisites¶

M12-W01 complete with target companies + format research.

Recommended cadence¶

Session A-Tue/Wed evening (~3 h): coding refresh + ML breadth
Session B-Sat morning (~3.5 h): ML system design + take-home prep
Session C-Sun afternoon (~2.5 h): behavioral stories + mock interview

Session A-Coding refresh + ML breadth¶

Goal: Refresh LeetCode-medium muscle. Self-quiz on ML breadth.

Part 1-Coding refresh (60 min)¶

If rusty, do 3-4 LeetCode-medium problems. Focus on: - Arrays + hash maps (most common in ML interviews). - Two pointers. - BFS / DFS. - Strings.

If comfortable, skip. Don't grind for grind's sake.

Part 2-ML breadth quiz (60 min)¶

Self-quiz, no notes: 1. Derive the gradient of softmax + cross-entropy. 2. Explain attention's √d_k scaling. 3. What does FlashAttention solve? 4. What's the difference between LoRA and full fine-tuning? 5. What's a KV cache and why does it grow with sequence length? 6. Differentiate fine-tuning vs RAG: when each? 7. Explain Cohen's kappa and when you use it. 8. What's BM25, conceptually? 9. What is bf16 vs fp16-why prefer bf16? 10. DPO vs PPO-why DPO?

For each weak answer: re-read your notes from the relevant Q1-Q3 week.

Part 3-Specialty depth (60 min)¶

For your specialty: 3 questions you'd expect to be asked. - (A) "How would you evaluate an agent's trajectory quality?" - (B) "How would you debug an agent that loops on the same tool?" - (C) "What's the bottleneck of LLM inference and how do you address it?"

Write 200-word answers for each. Practice saying them aloud.

Output of Session A¶

Coding refreshed.
Breadth quiz answers.
3 specialty Q&A's prepared.

Session B-ML system design + take-home prep¶

Goal: Practice 2 system-design problems. Set up a take-home environment.

Part 1-Read Chip Huyen's ML Interviews (60 min)¶

Read: Chip Huyen's ML Interviews Book-free at huyenchip.com/ml-interviews-book.

Focus on the system design chapter.

Part 2-Practice 2 ML system design problems (90 min)¶

Pick 2 from: 1. "Design a customer support agent for a SaaS company." 2. "Design a RAG system over 10M PDF pages." 3. "Design an LLM eval pipeline for prompt regression detection." 4. "Design an inference service for a 70B model serving 1000 QPS."

For each: spend 45 min sketching out loud (record yourself). Cover: - Clarifying questions. - Functional requirements. - Non-functional (scale, latency, budget). - Architecture diagram. - Data flow. - Tradeoffs. - Failure modes.

Compare to a "reference answer" you'd find on a blog or your own past work.

Part 3-Take-home prep (60 min)¶

If your targets do take-homes: prep an environment. - Template repo with your standard scaffolding (uv, pytest, ruff). - Quick LLM client + Pydantic + Anthropic / OpenAI imports. - A standard eval harness skeleton.

Goal: when a take-home arrives, you start at 80% setup-done.

Output of Session B¶

2 system-design recordings.
Take-home template repo.

Session C-Behavioral stories + mock interview¶

Goal: 5 STAR stories about your year. One mock interview.

Part 1-STAR stories (75 min)¶

5 stories, each ≤2 minutes spoken. Cover: 1. Capstone project-Situation, Task, Action, Result. 2. Bridge from SRE-why did you transition? 3. A failure-something that didn't work; what you learned. 4. Disagreement-when you pushed back on a technical decision. 5. OSS contribution-what you contributed; how it landed.

Each: speak aloud, time it, refine.

Part 2-Mock interview (45 min)¶

Options: - A friend in the field (best). - Pramp.com (free; pair-up algorithm). - Yourself recorded (worst, still useful).

60 minutes split: 15 min coding, 30 min ML system design, 15 min behavioral.

Part 3-Self-assessment (30 min)¶

What was weak? Where did you ramble? Which question stumped you?

Make a 5-item action list for next week's polish (M12-W03 has slack time).

Output of Session C¶

5 STAR stories rehearsed.
1 mock interview done.
Action list of weaknesses.

End-of-week artifact¶

End-of-week self-assessment¶

I can answer 8/10 ML breadth questions confidently.
I can sketch a RAG / agent / inference system in 30 min.
I can tell my year's story in under 2 minutes.

Common failure modes for this week¶

Over-grinding LeetCode. ML companies care less than you fear; refresh, don't grind.
No mock. The first time you say a story aloud, it sounds rough. Rehearse before the real interview.
Skipping system design. It's the hardest format and the one you can prep for most.

What's next (preview of M12-W03)¶

Year-2 plan + capstone v0.3 + year-in-review post draft.