In this article
Before you start
This page covers one round; how interview loops are structured shows where it sits among the others.
Why interviewers ask this
The project deep dive replaces an artificial exercise with real evidence. Ashby, a company that publishes its own guide for candidates, says most of its other interviews are simulations with artificial time limits and little context, while a past project shows how you gather information, make decisions, plan work and deal with setbacks. This is one company's stated reason, not an industry rule.
A weak answer lists features and says "we" throughout. A strong answer shows what you personally decided, what you weighed and what happened. Ashby also says that effective communication is an important skill it is assessing in this interview, so an answer the interviewer cannot follow can hurt even when the work was good. Hiring research supports asking every candidate job-related questions and judging the answers against set criteria: Schmidt and Hunter (1998) review decades of findings in which structured interviews predicted job performance better than unstructured ones. Other companies word the round differently; how interview loops are structured explains how rounds vary.
The project deep dive from first principles
- Project deep dive
- A project deep dive is an interview round in which a candidate explains one past project in detail, leading the walk-through while the interviewer interrupts with questions, so the interviewer can judge the candidate's decisions, ownership and communication from real work.
- Decision point
- A decision point is a moment in a project where you chose between at least two options, such as a database, an algorithm or a launch plan, and can explain what each option cost and why you picked one.
In plain words
Think of a travel diary read aloud to a friend who keeps asking "why did you take that road?" The diary is the project. The questions test whether you remember the map, not only the photos. The analogy breaks in one place. A friend wants a good story, while the interviewer wants evidence of your own judgment. So the story has to be about choices rather than scenery.
In one precise sentence, you present a project you led, mark its decision points, and defend each one with the options you rejected and the evidence you had.
Step 1: Choose a project you led and can explain
The best project is one where you made the critical decisions and saw the results. Ashby asks for a project you led and made critical decisions for, with some ambiguity, and one you can describe in both business context and technical architecture. Among the projects it is much less interested in is one that was partially released or that you left before it was released, because it is hard to know the outcomes of your decisions if you did not see them firsthand. That is Ashby's taste; ask the recruiter what the company prefers.
Input: your list of past projects. Output: one project with at least three decision points, of which your opening covers one or two and the rest come out under questioning. You now have the material; the next step turns it into answers.
Step 2: Defend each decision with the alternative you rejected
Each decision point needs four parts: the options, the one you chose, the cost you accepted and the evidence. Ashby says it wants you to explain why important decisions were made, including technology choices, tradeoffs, testing and reliability considerations, and to recall concrete details such as data-model properties, milestone timelines and customer reactions (all per its guide).
Say it in this order: "We could have done A or B. I chose B because of this evidence. It cost us this, and we handled it by that." The sentence about the cost is the one most candidates skip, and it is the one the follow-up question targets. If you cannot name a rejected option, you probably did not make that decision; pick another one.
Input: one decision point. Output: a four-part defense you can say in about fifteen seconds.
Step 3: Close with a result, a number and a lesson
The result tells the interviewer whether the decisions worked. Name what changed, measured by one metric, and say who noticed, such as users or the on-call team. If you have no number, say so and describe the observable change instead; an invented number is worse than none because follow-ups test it.
End with one lesson that changes how you work, stated as an action: "I now add a rollback switch before launch." A lesson such as "communication is key" gives the interviewer nothing to probe. Input: the outcome. Output: one metric and one changed habit.
Step 4: Fit the time and expect interruptions
The length depends on the round. Ashby says the first walk-through happens in an introductory call and takes about 10 to 20 minutes, and the final round runs 30 or 45 minutes depending on seniority. You may reuse the same project across rounds, but it needs enough depth to fill the longer slot.
You lead, and the interviewer stops you with questions along the way. Treat an interruption as a signal of what they want more of, answer it briefly, and return to your thread. Ashby says you are welcome to use diagrams or slides if they help you, and you decide where context is needed.
Step-by-step description
- Choose a project you led.
- Name the rejected option and its cost.
- State one result and one lesson.
- Fit the slot and expect questions.
- Choose a project you led and can explain.
- Defend each decision with the alternative you rejected.
- Close with a result and a lesson.
- Fit the time and expect interruptions.
Knowing it is half the job. Saying it is the other half.
Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.
Start Free Trial →The Staff-level answer
Walk me through a project you led and the decisions you made.
I led the move of our search ranking from nightly batch scoring to scoring at request time, because stale rankings left new items buried. We had two options: refresh the batch hourly, or score live. I chose live scoring and accepted higher serving cost, because even an hourly refresh left new items unranked for up to an hour, and a replay of one week of traffic showed the gap.
I first underestimated that cost, and I convinced the infrastructure lead by sharing the replay results. I held the latency budget by caching item features, not scores, so rankings stayed fresh. In the A/B test, click-through on new items went up and median latency held flat. Next time I would add a rollback switch from day one.
What makes it Staff-level: it names the rejected option, the cost accepted, the evidence that settled the choice, who was convinced and one change made in hindsight. The project and its figures are illustrative, not a real case.
The opening summary is the 60 to 90 seconds you speak before the first interruption: project, your role, the key decision, the result and the lesson. That range is an illustrative planning value, not a sourced figure; the round itself lasts the sourced length from Step 4. At an illustrative speaking pace of 2.5 words per second, a 90-second opening is 225 words, and the decisions get 35 of the 90 seconds, the largest share.1
Text description of the diagram
- The decisions get 35 of the 90 seconds, the largest share.
- At an illustrative pace of 2.5 words per second, 90 seconds is 225 words.
The rubric
This table is a self-check built from the sources on this page, not any company's scoring sheet. Interviewers differ, and no company's rubric is claimed here.
| Dimension | Weak answer | Strong answer | Staff answer |
|---|---|---|---|
| Ownership | Says "we" and cannot name their own part | States the decisions they personally made | Also says where they were wrong and who they convinced |
| Depth | Lists features and tools | Explains why each key choice was made | Recalls details such as data model, timeline and customer reaction |
| Trade-offs | Names no alternatives | Names one rejected option | Names the cost accepted and the evidence behind the choice |
| Communication | Rambles or over-explains basics | Follows a clear order | Adapts to interruptions and returns to the thread |
Three variants of this question
What were the three biggest challenges you faced in this project?
Pick three challenges that are each a decision point, not three complaints, and give each one its option, choice and cost. Order them by how much they changed the outcome, so that an interruption after the first still leaves your best material on the table.
Walk me through a recent large-scale project you worked on. How did you come to agreement with other teams?
Name the disagreement, the position of each side, the evidence you brought and what each side gave up. Candidates report this wording in loops for larger companies; the Meta engineering manager interview page shows how one such loop is organized. Keep your own role explicit, since agreement is rarely one person's work.
How did you break this project down into a science problem?
Candidates report this wording in applied-science and AI-lab loops. Answer by stating the business goal, the quantity you chose to predict or optimize, how you would know it worked, and what simpler baseline you tried first. The Amazon applied scientist interview process page covers the loop around such questions.
Common mistakes
- "We built a recommendation system." (wrong) The interviewer cannot tell what you did. (right) "I owned the ranking model and made the choice between two designs."
- "I picked the biggest project I ever touched." (wrong) Size does not matter if you cannot defend the decisions. (right) "I picked the project where I led the decisions and saw the result."
- "I prepared slides covering every detail of the project." (wrong) The error is length: there is no time to show it all, and the interviewer interrupts. (right) "I brought one diagram of the architecture and left time for questions."
- "The result was great, no downsides." (wrong) It invites the follow-up about costs and failures. (right) "It worked, and the cost was this; I would change that next time."
Follow-up questions
Do you need slides for a project deep dive?
Not necessarily. Ashby says you are welcome to use diagrams or slides if they help you. One clear architecture diagram is usually enough, and your own company's instructions come first.
Can you use the same project in every round?
At Ashby, yes, provided the project has enough depth to fill the longer final round. Elsewhere, ask the recruiter, and keep a second project ready in case the interviewer already heard the first.
What if the project was never released?
Ashby lists projects you left before release among the ones it is less interested in, because you could not see the outcome of your decisions firsthand. Prefer a project with a visible result; if all you have is an unreleased one, say clearly what you learned from its design and what you could not measure.
How long should a project deep dive answer be?
The round lasts the length you were told; Ashby's is about 10 to 20 minutes in the first call and 30 or 45 minutes in the final round. Your uninterrupted opening is about 90 seconds (an illustrative planning value), with the decisions taking the largest share, and the rest of the time is the interviewer's questions. Ask the recruiter about the format. The AI engineer interview questions page gives more questions to rehearse around the same project.
What should you do when the interviewer interrupts?
Answer the question in a sentence or two, then say where you were and continue. The interruption shows what the interviewer wants more of. The OpenAI interview questions and Anthropic interview process pages show how rounds in AI labs are laid out, and the forward deployed and applied AI engineer rounds page covers roles where project discussion is central.
- The project deep dive is one real project you led, explained while the interviewer interrupts.
- Choose a project where you made the critical decisions and saw the result.
- Defend every decision with the rejected option, the cost accepted and the evidence.
- Close with one metric, or an honest description of the change, and one lesson stated as an action.
- Slides are optional; plan the time for the round format you were told.
Sources
- Engineer Past Projects Deep Dive — Ashby (candidate interview guide; accessed 2026-10-26), 25 Jul 2025
- The validity and utility of selection methods in personnel psychology: practical and theoretical implications of 85 years of research findings — Psychological Bulletin (Schmidt and Hunter), 1 Sep 1998
