Final Exam timer (30 minutes for 15 questions): 00:00
Start 30:00 exam
Rules: no scrolling back to other pages. Answer honestly — this tells you what to fix in your last hour. Pass = 70% (11/15).
Test yourself
Final Exam — All Modules Mixed Tap an answer — explanation appears instantly. Pass mark: 70%. Retake until you pass.
An LLM pipeline must output data your code consumes. Safest design?
Free text, parse with regex and hope JSON schema + validation + one retry with error feedback + human fallback Temperature 1 for variety Longer prompts only Why: Schema, validate, retry, fallback. The production pattern from Module 1.
Webhook endpoint does heavy LLM work before replying. Problem?
None — thorough is good Provider times out and retries; you'll process duplicates JSON gets prettier Costs drop Why: Return 200 fast, enqueue, process async.
First move on an intermittent workflow failure?
Rewrite everything Reproduce with smallest failing input; check logs/correlation id Blame the provider publicly Add retries everywhere blindly Why: Reproduce → isolate → fix → verify → document.
User wants answers from 500 internal PDFs, updated weekly. Design?
Fine-tune weekly RAG: chunk, embed, retrieve top-k, cite sources Paste PDFs into the system prompt Keyword search only, no LLM Why: Changing private knowledge = RAG.
Claude Code suggests a large refactor mid-bugfix. You…
Accept — momentum Keep scope small: fix the bug with a test first; refactor separately with review Paste prod secrets for context Turn off tests to go faster Why: Small scoped changes, reviewed diffs, tests first.
429s during a batch job. Best combined fix?
Parallelise harder Backoff + jitter + queue + cache + respect Retry-After Retry every 100ms Rotate API keys to evade limits Why: Be a good API citizen; queue and cache.
Confidence from the classifier is low on 15% of tickets. Design?
Auto-send anyway Route to human review queue; log for eval set improvement Discard them Raise temperature Why: Human gate for low confidence; feed corrections back into evals.
Best proof you can write SOPs/documentation?
Saying 'I document well' Show one SOP: purpose, trigger, steps, failure modes, escalation, changelog A 200-page dump Screenshots only Why: Structure proves documentation skill.
Embedding similarity is used for…
Hashing passwords Semantic search / RAG retrieval / clustering Compressing images Signing JWTs Why: Meaning as vectors.
Your prompt change 'feels better'. Ship it?
Yes, ship fast Run the golden set; ship only if accuracy/cost improve Ask the model if it's better Ship on Fridays Why: Evals over vibes.
Dedupe table / idempotency key prevents…
Slow queries Duplicate records when providers or clients retry Prompt injection fully Rate limits Why: Safe retries without duplicates.
STAR answer is failing in mocks. Likely cause?
Too much Result All Situation, no Action/Result — rebalance to 60% Action with a measurable outcome Speaking too clearly Using numbers Why: Action + measurable Result is what gets hired.
New AI tool evaluation for the team — deliverable?
A hype post Spike results + golden-set eval + pricing/limits/privacy/lock-in trade-offs + recommendation Immediate migration A longer prompt Why: Spike, eval, trade-offs, recommendation.
API keys were committed to git. Immediate steps?
Delete the last commit message Revoke/rotate the keys, remove from history, move to env/secrets, add .gitignore + scanning Ignore if repo is small Rename the file Why: Committed = leaked. Rotate first.
Interviewer: 'Design reporting for classified tickets.' Strongest close?
'Use a dashboard' Daily/weekly report by category + accuracy & cost metrics + low-confidence review stats, with owners and alerts 'Email everything to everyone' 'The LLM will handle it' Why: Reporting + quality + cost + ownership closes the system-design answer.
0 / 15 correct