START HERE · NO PREREQUISITES
Navigate the ML design interview.
Explain a complete decision system, then deepen the parts that matter for the role. You do not need forty theory lessons before trying a small design.
Confirm the role and round
Ask the recruiter what the round assesses. Applied ML emphasizes product decisions, data, representations, models and evaluation. ML infrastructure emphasizes platform contracts, resources, reliability and operation. Research and research engineering need specialist preparation beyond this curriculum. Do not infer interview format from a company logo.
Choose an entry path
Applied ML: framing → events → labels → leakage → classification metrics. Then practice the complete Video Recommendations and Content Moderation designs with their supporting foundations.
Platform: events → labels → leakage → metrics, with framing as needed. In a platform design, center temporal correctness, contracts, compatibility, isolation, rollout and recovery. Feature/training/serving platform chapters are planned.
Existing introductory lessons remain available in the sidebar. Advanced MLF-040 will be a cumulative capstone; this orientation provides the early framework without its advanced prerequisites.
A 45-minute rehearsal
This is an original exercise budget, not an employer’s official interview format. Adapt it to the time available and interviewer direction. The minimum evidence column is a stopping point, not a script to recite.
| Minutes | Purpose | Minimum evidence |
|---|---|---|
| 0–5 | Frame the decision | Actor, action, prediction unit/horizon, constraints, product outcome and one harmful failure. |
| 5–8 | Establish a complete baseline | A usable simple policy and an input → decision → feedback lifecycle. |
| 8–18 | Design data and labels | Availability, target definition/delay, split rule and prioritized signal groups. |
| 18–28 | Choose and explain a model | One justified upgrade, input/output shapes, objective, sampling and rejected alternative. |
| 28–36 | Evaluate and integrate | Offline denominator, online outcome/guardrail, serving boundary, fallback and release evidence. |
| 36–42 | Investigate a high-risk trade-off | Change an assumption, trace the consequence and describe verification. |
| 42–45 | Close the loop | Remaining risks, deferred decisions, measurement and the next experiment. |
For a platform round, redirect the model interval toward workload, state, correctness and resource isolation. For an applied round, keep detailed storage internals as follow-ups while still explaining availability and compatible serving. The full-study chapters retain the deeper schema and failure analysis.
Recover from an interruption
Answer the probe, state which assumption changed, revise the affected decision, then return to the nearest unfinished goal. Keep a short deferred-topics list. Example: “That changes which labels are available; I would revise the split and quality gate first, then return to serving.” When time is short, offer two relevant deep dives and follow the interviewer’s choice.
Discuss signals through hypotheses
Choose a few useful signal groups. For each, state the source, why it may predict the target, representation/shape, availability window, missing default, freshness/cost and an ablation or failure test. Three groups plus a rejected signal is a rehearsal default, not a universal count.
For a recommendation request, consider eligible item metadata, prior user history and request context. Reject watch completion from the same request as an input; it is a possible later outcome. “Use embeddings” is incomplete: explain training signal, score direction, negative sampling, new-entity behavior and compatible index version.
Diagnose before adding complexity
Low training loss and high validation loss suggest several hypotheses: overfitting, different populations, leakage repairs or transform mismatch. High losses on both may reflect insufficient signals, optimization or label quality. A later-time regression may reflect drift, changed policy or serving skew. Name a competing hypothesis and a discriminating check before choosing regularization, new data or another model.
Compare a rule/simple model, a justified upgrade and an optional research-informed alternative. Explain what evidence would keep the simpler solution. A dated research card should identify the transferred idea, assumptions and limits; a toy experiment does not reproduce a large production benchmark.
Debrief with observable evidence
| Evidence state | What you can demonstrate |
|---|---|
| Missing | Cannot yet define the action, relevant data or metric. |
| Prompted | Can explain after a hint, but misses a consequence. |
| Independent | Justifies a coherent design and checks a boundary without prompting. |
| Adaptive | Revises the design under a changed constraint and traces consequences. |
Assess framing, data, baseline/model, evaluation, integration, operation and communication separately. Engineer evidence includes correct examples and metrics; Senior evidence adds rollout/failure containment; Staff evidence adds ownership, objective conflict and migration. These are teaching lenses, not an official hiring rubric.
After each mock record one demonstrated strength, two missing pieces of evidence, a targeted next exercise and a retest question. Reading progress is separate from readiness. No score predicts an offer.
Practice the first complete system
Try Video Recommendations → · Read the complete worked design →
Try Content Moderation → · Read the moderation worked design →
Try Feature Platform → · Read the feature correctness worked design →
Start the first detailed chapter
Frame the decision and non-ML baseline →Original preparation guidance. Source background: Google Rules of ML. Synthetic examples; no model training or automatic evaluation. Author self-review; independent technical review pending.