Week 12 Checkpoint + Interview Drill II
- Recall Week 12's core ideas — clustering, PCA, error analysis, tracking, pipelines — from memory
- Answer 20 rapid-fire ML-fundamentals questions aloud, interview-style, and self-grade against model answers
- Present the churn project in a tight written walkthrough: problem → approach → result → next steps
- Update the interview error log with this phase's weak spots and schedule their revisits
| Spaced-rep: full Week 12 due deck (SM-2 queue) | 15 min |
| Guided: closed-book Week 12 recall + diff | 20 min |
| Guided: 20-question mock, aloud and timed | 40 min |
| Practice: model answers for the weakest five | 15 min |
| Project: walkthrough + error-log update, commit | 20 min |
| Cumulative quiz + retag weak flashcards | 10 min |
Builds on: Day 83 — Churn project (the thing you will defend) · Day 77 — Week 11 checkpoint & model card · Day 28 — The interview error log
In medical school, learning ends with a viva: you stand before examiners and explain — out loud, without notes — what you did for the patient and why. Not because talking is medicine, but because fluent explanation is the sharpest test of understanding. If you can only recognize an idea when someone else says it, you have familiarity. If you can produce it cold, defend it against a "but why not…?" and attach a number from your own work, you have knowledge.
Today is the ML viva. Two hours, three movements: drag Week 12 back from memory before it fades (that struggle is the spaced-repetition rep that makes it stick); face 20 rapid-fire questions of the kind real screens actually ask — "precision vs recall?", "your model overfits, name three fixes", "trees vs linear, when and why?" — answering aloud, at interview tempo; and write the five-minute walkthrough of yesterday's churn project, because "tell me about a project" is the one question you are guaranteed to get, and the difference between rambling and a crisp problem → approach → result → next-steps arc is rehearsal. The error log from Day 28 gets its update: every stumble today is scheduled for a revisit, which is how weak spots become former weak spots.
Phase 4 ends here, and ML fundamentals are the most predictable screen in the hiring pipeline: recruiter phone screens, take-home debriefs, and the "ML basics" round all draw from the same ~40-question well you drill today. AI-engineer roles have not dropped these questions — they ask them right before the LLM questions. The project walkthrough matters even more: interviewers weight "walk me through something you built" far above trivia, and your churn repo plus today's rehearsed narrative is a complete, honest answer. The error log habit — log, revisit, retire — is the compounding engine behind Day 179's final interview gym.
Guided practice
Week 12 recall, closed book
20 minNotes shut. On paper, from memory:
- The k-means loop, and its three built-in assumptions. (Day 78)
- Silhouette score: formula shape and how to read +1 / 0 / negative. (Day 78)
- PCA's objective in one line, and the two-word reason a 2-D overlap plot does not prove classes are inseparable. (Day 79)
- The error-analysis loop: what you look at first, second, third after a model underperforms. (Day 80)
- What an MLflow run records, and the three things a fully reproducible run pins. (Day 81)
- The predict-function contract from the pipeline template: inputs validated, outputs shaped how? (Day 82)
- The churn project's leakage feature and the audit question that caught it. (Day 83)
Open notes, diff, mark ✓/~/✗. Re-derive every ✗ from its source day now — the viva assumes them.
The 20-question mock, aloud
40 minSet a timer: 90 seconds per question, answer ALOUD, then check the model-answer notes you will write afterward. Score each 0/1/2 (blank / partial / full three-part answer).
- Accuracy is 94% — when is that meaningless, and what do you ask for instead?
- Precision vs recall: definitions, denominators, and one business that optimizes each.
- ROC-AUC vs PR-AUC — which for 1% fraud, and why?
- Your model: train 0.98, validation 0.72. Diagnose and give three fixes.
- Your model: train 0.61, validation 0.60. Diagnose — will more data help?
- L1 vs L2 regularization: mechanical difference and when you want each.
- Why must the scaler live inside the CV pipeline?
- What is the winner's curse in hyperparameter search, and the defense?
- Trees vs logistic regression: three deciding factors.
- Bagging vs boosting in two sentences, including which variance/bias each attacks.
- Why do random forests not overfit as trees are added, while boosting rounds can?
- Feature importances: two reasons to distrust them, and the honest alternative.
- k-means: the loop, and why it always converges but not always well.
- How do you choose k? Name two methods and their weaknesses.
- When does DBSCAN beat k-means? What are its two knobs?
- PCA in one line — and why standardize first?
- Explain the curse of dimensionality's effect on nearest-neighbor methods.
- What is data leakage? Give two distinct examples and the audit question.
- A stakeholder asks "is the model good?" — walk your answer structure.
- What does "reproducible" mean for an ML experiment? Name what you pin.
Total /40. Under 30: schedule the weak families' source days for revisit this week.
On your own
Write the model answers
15 minFor your five LOWEST-scoring questions from the mock, write full-credit model answers using the three-part structure (direct answer → mechanism → example-with-number from your own work). Then answer each aloud once more, from memory, at tempo.
Constraint: every model answer must cite a number or artifact from YOUR Week 11–12 repos (the depth-sweep table, the silhouette sweep, the churn precision@k at the retention budget, the leakage mirage score from the curiosity checkpoint…). Generic answers are the thing this practice exists to kill.
Hint: if a question has no number anywhere in your repos, that is itself a finding — go generate it in ten minutes or log it as a gap.
Viva packet: walkthrough + error log update
Ship week12_viva/ in your practice repo: (1) WALKTHROUGH.md — the four-beat churn presentation (problem, approach incl. the leakage story, result with caveat, next steps), ≤ 350 words, rehearsed aloud at least twice, plus a short appendix defending two of your Day 83 self-grade scores with evidence; (2) mock_answers.md — your 20 scores plus full model answers for the five weakest; (3) an updated interview error log (from Day 28): add every 0–1-scored question with its family, source day, and a scheduled revisit date; retire any old entries you now answer cold. Commit all three — Day 179's interview gym starts from exactly this log.
Common mistakes & misconceptions
- Reviewing by rereading notes. Recognition masquerades as recall; only closed-book production and aloud answers reveal what interview pressure will reveal anyway.
- Answering rapid-fire questions with definitions only. Full credit = answer + mechanism + example; the example from your own project is what interviewers remember.
- Telling the project story chronologically ("first I loaded the data, then I…"). Lead with the problem and the result; the four-beat arc beats the diary every time.
- Hiding the leakage story or the CV-to-holdout drop in the walkthrough. Honest caveats are credibility signals — interviewers probe for exactly whether you know your model's weaknesses.
- Logging errors without scheduling revisits. An error log that is never reread is a diary; the revisit date is what turns it into a training plan.
- Treating today as optional because Week 12 "just happened." The forgetting curve is steepest in the first days — this session is timed to catch memories at peak leverage.
Q1. A silhouette sweep peaks at k=3 but the elbow looks like k=5, and the business swears there are 4 segments. What is the defensible move?
Q2. In your churn walkthrough, the strongest single credibility signal to include is…
Q3. PCA reduces your 64 features to 10 components and your tree ensemble's PR-AUC drops slightly. The likely explanation and right response is…
Go deeper — curated resources
- LLM Interview Questions repo — preview the next phases' bank ↗ — Skim the fundamentals section: notice how many questions are Phase 4 concepts (metrics, overfitting, evaluation) wearing LLM clothes. You will drill these properly from Phase 6 on.
- Week 12 recall drills diffed; every ✗ re-derived from its source day
- Mock scored /40 with family totals; weak families scheduled for revisit
- WALKTHROUGH.md rehearsed aloud twice and committed
- Error log updated with dated revisits; retired entries pruned
- Cumulative quiz ≥ 2/3
← Back: The viva defends Day 83's build using Week 11's vocabulary (Days 71–77) and Week 12's craft (Days 78–82). The error log you updated was born on Day 28, and the aloud-answer discipline started at Day 35's first interview drill.
Forward →: Phase 5 begins tomorrow: Day 85's neural networks reuse function+loss+optimizer verbatim, Day 89 reads loss curves with today's bias/variance eyes, and Day 90's MNIST lab expects the Day 82 template you just defended. The walkthrough and error log feed directly into Day 179's interview gym and Day 180's demo day.
Unlocks: D179 Interview Gym