09 · Interview preparation
A framework for the 45 minutes
12 min read
A framework for the ML design interview is a fixed order of six phases with a number of minutes attached to each. It exists so that a 45-minute round covers framing, data, the model, evaluation, serving and monitoring without running out of time. It changes one design decision above all others, which is where you stop talking about one part and move to the next. This topic gives the six phases, a clock for them, a spoken opening you can adapt, and the habits that keep you on time.
An everyday example
A museum guide has 45 minutes to take a group of eight through a building with twelve rooms. A good guide has a route, spends the longest in the two rooms the group came for, and says one sentence at every doorway so the group knows the room is finished. When someone asks a long question about a painting in a room the guide planned to pass through, the guide answers briefly and says they will come back to it at the end if there is time.
A poor guide starts in the first room, talks for twenty minutes because the first room is the one they know best, and then rushes the group past the rooms they paid to see. The interview framework is the guide’s route. The phases are the rooms, the minutes are how long the group stays in each, and the sentence at the doorway is the transition that tells the interviewer where you are.
A museum guide decides the route alone, while in an interview the person asking questions can send you to a different room at any moment, and there the comparison stops matching, so the framework has to bend when they do. The sections on reordering and clock habits are about bending without losing the route.
Six phases and a clock for a 45-minute round
The six phases are framing and scope, a pipeline sketch, data and features, modelling, evaluation and serving, and monitoring with deep dives. Framing takes five to seven minutes, the sketch three to five, data and features ten, modelling ten, evaluation and serving eight, and monitoring and deep dives take whatever is left, which is usually five to nine minutes. The order matches how a system is built, because you cannot choose features before you know the prediction and cannot choose a metric before you know the features.
The budget is uneven on purpose, with data and modelling getting ten minutes each because that is where the decisions that distinguish candidates are made, and the sketch getting three because its job is only to give the later phases a picture to point at. If you find yourself twelve minutes into the sketch, the extra minutes have come out of data or evaluation, which is where the grade is decided.
The timeline below marks where the clock should read when each group of phases begins. The marks are targets for a rehearsal, and a real interviewer will push some phases longer with questions. What matters is noticing the gap between the clock and the target, which is the signal to speed up or to say that a detail will wait.
| Phase | Minutes | What the interviewer wants to hear |
|---|---|---|
| Framing and scope | 5 to 7 | Who the user is, what action the prediction changes, the input and the target, two or three numbers about scale, and the deadline. |
| Pipeline sketch | 3 to 5 | Five or six boxes showing how a request gets a prediction and how outcomes get back into training, with no component chosen yet. |
| Data and features | 10 | Where labelled examples come from, how rare the positive case is, and a few signals per source with a reason for each. |
| Modelling | 10 | A baseline, one stronger alternative, the scoring rule it trains with, called the loss, and why the data volume and the deadline allow it. |
| Evaluation and serving | 8 | An offline metric tied to a product metric, the threshold or ranking rule, and how the prediction fits inside the deadline. |
| Monitoring and deep dives | The rest | How you would notice the model getting worse, when you would retrain, and one or two of the interviewer’s chosen follow-ups. |
No phase passes ten minutes unless you are asked
The single rule that protects the round is that no phase runs past ten minutes unless the interviewer asks for more. Ten minutes is the budget of the two largest phases, so any phase that reaches it has used its full share, and the only good reason to stay is that the interviewer has said they want to go deeper here. If they have not said so, finish the sentence you are on, say what you have decided, and move.
The rule is easy to break in the modelling phase, because models are the part most candidates enjoy. A candidate who spends eighteen minutes comparing architectures has left eight minutes for evaluation, serving and monitoring together, and those three phases carry two of the five grading areas. A model described in ten minutes with a metric and a serving plan beats a better model described in eighteen minutes with neither.
When you reach ten minutes in a phase and still have more to say, the move is to earmark. Earmarking means naming the detail in one sentence and saying it is available in the deep dive, for example that you would be happy to come back to the choice of loss if useful. The interviewer now knows you have the depth and can ask for it when they choose, and you have kept your minutes.
A sample 90-second opening
The opening below is for a content moderation prompt, and it covers what a listener needs in the first minute and a half. It states the user and the action, asks the two or three questions that change the design, offers a task statement for the interviewer to accept or correct, gives two numbers and names the order the rest of the conversation will take. Read it aloud once with a clock, then rewrite it in your own words.
“Let me make sure I have the problem. Every video uploaded to the app has to be checked before it goes live, the check has to catch content that breaks the rules, and the content can be in the picture, the sound or the caption, so I need to look at all three together. Before I draw anything I want to ask about three things. I need roughly how many uploads arrive each day, so I can tell whether the check must be cheap. I need to know how long an upload can wait before it is published, because that decides whether a human can be in the loop. And I need to know whether there is already a team of reviewers, because a design that sends them more than they can read is a design that fails.”
“If I assume 100 million uploads a day and a wait of up to a few minutes, that is about 1,200 uploads a second, so the first stage has to be cheap and automatic. My task statement is a model that gives each upload a probability of breaking the rules, where a very high probability removes the upload, a very low one publishes it, and the middle band goes to a reviewer. If that framing works for you, I will sketch the pipeline in a couple of minutes. Then I will spend most of the time on where the labelled examples come from and on how I would set the two thresholds, because with harmful content under one in a thousand uploads those are the hard parts.”
Two things make this opening work. Every number is labelled as an assumption the interviewer can correct, and the final sentence announces where the time will go, which gives the interviewer a chance to redirect you before you have spent anything.
When and how to reorder the phases
The default order assumes the hardest part of the problem is the data and the model. When a different constraint dominates, the phase that handles it moves forward, and you announce the move in one sentence so the interviewer knows it is a choice. The two common cases are a problem where the deadline dominates and a problem where judging the output dominates.
A deadline-dominated problem is one where the response time rules out most designs before any model is chosen. Search autocomplete, which must return completions within about 100 milliseconds of a keystroke, and fraud scoring, which must answer before a payment completes, are examples. For these, pull serving forward to sit directly after the sketch, say that the deadline shapes everything downstream, and only then discuss data and the model within the constraint the deadline sets.
An evaluation-dominated problem is one where the hard part is knowing whether an output is good at all. A system that writes answers to customer questions has no single correct output to compare against, so pull evaluation forward and agree on how outputs will be judged before discussing how to produce them. The announcement is short, for example that because there is no single right answer here you want to settle how you will grade outputs before designing the model that produces them.
Habits that keep the clock honest
Three habits keep a candidate on the route without seeming rushed. The first is the one-sentence transition, which closes a phase by stating what was decided and names the next one, for example that with the task and the deadline agreed you will now sketch the pipeline. The sentence tells the interviewer where you are and gives them a natural place to redirect you.
The second habit is earmarking instead of diving. When a detail tempts you, say it exists and that you will return to it, then continue. A candidate who dives into every interesting detail finishes the round with a deep answer on one phase and nothing on three others, while a candidate who earmarks finishes with a complete design and a list of deep dives the interviewer can choose from.
The third habit is recovering without apology. If you notice at minute 25 that you are still on data, do not apologise or explain, because both cost time and draw attention to the slip. Say the one-sentence transition and move to modelling with a slightly shorter plan for it. Interviewers remember a candidate who adjusted calmly far better than one who said sorry three times.
Where people get confused
Newcomers treat the minute budget as a script to recite, and then feel lost when the interviewer asks a question that belongs to a later phase. The budget is a route, and a question from a later room is answered briefly where you stand, with a sentence saying you will cover the rest when you reach that phase. The interviewer is allowed to move you, and the framework only needs you to know where you are afterwards.
The sketch phase is also mistaken for the place where the design is decided, so candidates draw for ten minutes. The sketch exists so that the later phases have a picture to point at, and five boxes drawn in three minutes do that job. Every decision that matters, from the label to the threshold, is made in the phases after the sketch.
The deep dive phase is taken for optional padding at the end. It is where the interviewer tests depth on the one or two parts they care about, and a candidate who has earmarked well arrives there with a list to offer. Arriving with no earmarks and no time is the most common way a strong design earns a weak grade.
In the interview
The framework is the whole 45 minutes, so there is no separate moment to bring it up. Announce it once in the last sentence of your opening, by saying the order you will take and where the time will go, and then let the transitions carry it. Three warning signs, each with a recovery, cover most of the ways a round goes wrong.
| Warning sign | How it feels in the room | Recovery |
|---|---|---|
| Fifteen minutes gone and nothing is drawn | You are still asking clarifying questions or describing the product. | Say that you have enough to start, state the task in one sentence with one number, and draw five boxes in two minutes. |
| The interviewer asks the same question twice | Your first answer went around the question rather than through it. | Answer the question directly in one sentence, then give the reason. If you do not know, say so and describe how you would find out. |
| Ten minutes with no number spoken | You are listing components and adjectives such as fast and scalable. | Stop and do one piece of arithmetic aloud, such as daily volume divided by 86,400 seconds, and say what the result forces the design to do. |
Interview questions
The prompt is fraud detection for a payments company. Walk me through how you would spend the 45 minutes.
I would spend about six minutes on framing, confirming that every payment needs a score before it is approved and that the answer must come back within about 50 milliseconds, and I would state the task as predicting the probability that a payment is fraudulent. Because that deadline rules out most designs on its own, I would announce that I am pulling serving forward to sit right after the sketch.
The sketch would take three minutes and serving about six. Then I would give ten minutes each to the data, where the labels arrive weeks late when a chargeback comes in, and to the model, starting from a baseline of rules and moving to a tree-based model, which sorts payments with a series of yes-or-no questions. That leaves roughly ten minutes for evaluation, where I would report how many flagged payments are real fraud at a fixed share of frauds caught, because every blocked legitimate payment costs a customer, and for monitoring and whatever the interviewer wants to go deeper on.
You are at minute 28 and still discussing features. What do you say?
I would finish the sentence I am on, then say that I have the main signals and will name any others if they become relevant, and that I am moving to the model now. I would not apologise, because an apology costs time and draws attention to the slip, and the interviewer mostly notices whether I adjust.
I would then shorten my modelling plan to a baseline, one alternative and the loss, in about six minutes, so that evaluation and serving still get time. If the interviewer wanted more on features, they would stop me, and that would be their choice of deep dive rather than my overrun.
The interviewer asks about monitoring while you are still framing the problem. How do you handle that?
I would answer the question briefly where I stand, for example by saying that I would watch the rate at which the model flags items against the rate reviewers confirm, and that a gap opening between them would be my signal to retrain. That takes about thirty seconds and shows I have thought about it.
Then I would say that I will come back to monitoring properly once the task and the data are settled, since the right things to watch depend on what the model predicts, and I would return to framing. Reported senior rounds at large companies press on exactly this kind of jump, so being able to answer and then return is worth practising.