In this article
Before you start
The worked example is a search project from the embeddings and vector search field, so two search terms appear. The advice itself applies to any project.
Why interviewers ask this
A vague answer here costs you the one moment of the interview where you control the story. The question is open, so the interviewer learns what you choose to tell: which project, which part was yours, and whether you can show an effect.
Amherst College describes the STAR method as a way to highlight action and impact. Google Operations Center's hiring guidance advises structured responses with data, so the listener sees how big your impact was. Wikipedia summarizes STAR as a structured technique employers use for behavioral and situational questions.
A weak answer lists tools and says "we built". A strong answer picks one project, says what you personally decided and gives a result with a number. A Staff-level answer adds the one thing you would change.
Past-project answers from first principles
- Trade-off sentence
- A trade-off sentence is a single sentence in a project answer that names the option you set aside and the reason you chose another, so the interviewer hears your judgment and not only your actions. It usually sits in the part of the answer about what you did.
- Project answer
- A project answer is a short spoken account of one specific piece of work you did, told as setting, task, your own actions and a measured result, with most of the time spent on what you personally decided and built.
In plain words
Think of a short news report about something you built. It gives one line of background. Most of the report is what happened. It ends with the outcome. A good report does not list every tool or every meeting; it follows one thread from the problem to the result, and the listener can repeat the story afterwards. The analogy breaks in one place: a reporter stays neutral and says "they", while you must say "I" and own your decisions.
In one precise sentence, a project answer is one specific event told in four parts, with most of the time spent on your own actions.
Step 1: Pick one specific project
A project answer starts with choosing one specific project, not with a tour of your resume. The University of Iowa says to describe a specific event, not a general description of what you have done. Google Operations Center's guidance also says to re-read your resume and the job description and find where they connect.
Input: your resume and the job description. Output: one project. A good pick has three traits. You owned a decision in it. It produced a result you can measure. It touches something the job asks for. For a search or retrieval role that might be a project using embeddings; for another role it is a different project.
You now have the subject of the answer. The next steps decide how long to spend on each part of it.
Step 2: Set the scene in about a third of the time
The Situation and the Task together set the scene, and they get the smallest share of the answer. Massachusetts Institute of Technology (MIT) career advising gives the Situation 20% and the Task 10%, and says the interviewer does not need every detail.
For the total, University College Dublin (UCD) recommends a response time between 90 seconds and 2 minutes. This page uses 90 seconds, the short end of that range. At the MIT shares the Situation takes 18 seconds and the Task 9 seconds; at 2 minutes they take 24 and 12 seconds.1
Input: the project. Output: two or three sentences covering where it happened, what was wrong, and what you were responsible for. Cut anything the Action does not need, such as team names, dates and tool lists.
Step 3: Say what you did, with I and one trade-off
The Action is where the answer is won: MIT gives it 60% of the time, which is 54 seconds of a 90-second clip.1 Say "I" for your own steps. The University of Iowa says to keep the focus on you and describe what you actually did, not what the team did.
Then add one trade-off sentence. A trade-off sentence is one sentence that names what you chose, what you gave up and why. For example: "I added a reranker rather than switching to a larger embedding model, because I could test it alone and it cost less to run."
This sentence is the rubric line from the lead. It shows judgment, not only effort, and it is the sentence an interviewer can ask about next. Input: your steps. Output: three or four "I" sentences and one trade-off sentence.
Step 4: End with a measured result
The Result closes the answer and gets the remaining tenth after MIT's 20, 10 and 60 percent, 9 seconds of a 90-second clip.1 Google Operations Center advises using data whenever possible, so the listener sees how big your impact was. The University of Iowa notes that students often must be prompted for results, so include them unasked.
A measured result has three parts: the number, the baseline it is compared with, and one honest cost. "Recall at 10 rose from 0.62 to 0.78 (illustrative), and tail latency rose too" is a result. "It went well" is not.
Input: your before and after figures. Output: one sentence with the change and its cost. The diagram below shows the whole answer on one timeline.
Step-by-step description
- Choose one specific project.
- Situation and Task take 27 seconds.
- Action takes 54 seconds.
- Result closes in 9 seconds.
- In a 90-second answer, Action gets the most time: 54 seconds, against 18 for Situation, 9 for Task and 9 for Result.
- The Situation takes 18 seconds and the Task 9 seconds.
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
Can you describe your past projects?
Let me pick one: a search system for support articles that I owned end to end. Embedding search alone left too many relevant articles outside the top ten: recall at 10 was 0.62 on about two thousand labeled queries. My task was to raise it without a slow page. I added a reranker rather than a bigger embedding model, because I could test that one change alone. I built the labeled set, ran the comparison and kept the reranker. Recall at 10 rose to 0.78, and tail latency rose from 120 to 180 milliseconds, which the team accepted. Next time I would rerank only the top 50 candidates.
All figures in this sample are illustrative. In the illustrative project, recall at 10 rose by 0.16 in absolute terms, a relative gain of 25.8%, while tail latency rose by 60 milliseconds, or 50%.1
Text description of the diagram
- Illustrative recall at 10 rose from 0.62 to 0.78 in absolute terms, by 0.16.
- In the illustrative project, recall at 10 rose by 0.16 in absolute terms, a relative gain of 25.8%.
What makes it Staff-level: the answer names a rejected option, states a cost next to the gain, and ends with a change. The model for the reranker is described in the cross-encoder reranker page. Compare a weak version of the same project, also illustrative: "We built a search system with embeddings. It worked well and everyone was happy." The strong version differs by the baseline of 0.62 recall at 10 (illustrative) and the cost it states.
The rubric
| Dimension | Weak answer | Strong answer | Staff answer |
|---|---|---|---|
| Focus | Lists several projects and tools | One specific project, chosen to fit the job | One project, tied to the role's needs in a sentence |
| Ownership | "We built it" | "I" for each of your own steps | Separates your decisions from the team's work |
| Result | "It went well" | A number with a baseline | A number, its cost and what it meant |
| Trade-offs | None mentioned | One option rejected, with a reason | The reason, and the evidence you used to choose |
| Follow-up readiness | Cannot go deeper | Can explain any step you named | Volunteers what you would change |
This table is advice about what a strong answer shows, not a description of any company's scoring.
Three variants of this question
Walk me through a project on your resume
The same shape answers "tell me about a project you are most proud of": pick the project where you made a decision and can show a number. Start from the resume line. Name the project in one phrase, give the context in one or two sentences, and spend most of the answer on your own actions and the result. Choose the project that best matches the role.
Which part of the project was yours?
Say what you personally decided and built, using "I", and name the teammates' work separately in one clause. The question tests the ownership row of the rubric, so give two or three concrete actions rather than a title.
What would you change if you did it again?
Name one specific change and the reason. For the sample answer above, reranking only the top 50 candidates, or distilling the reranker into a smaller model, would cut the added tail latency, because the cost comes from scoring every query and document pair. A change tied to a measured cost shows you learned something from the result.
Common mistakes
- "I've worked on a lot of things, such as search, ranking and some dashboards." (wrong) It is a list, not an event. (right) "One project I can walk through in detail is a search system I owned."
- "We built a retrieval pipeline and we improved it." (wrong) The interviewer cannot tell what you did. (right) "I built the labeled query set and I ran the comparison."
- "It went really well." (wrong) It gives no measurable outcome. (right) "Recall at 10 rose from 0.62 to 0.78 (illustrative), and tail latency rose to 180 milliseconds."
- Spending most of the time on background. (wrong) UCD names this the most frequent mistake, up to 70% of the time on background, while the Situation and Task together should get about a third of the answer and the Action most of it. (right) Two or three scene-setting sentences, then your steps.
- Treating a percentage split as an official rule. (wrong) Sources differ. MIT gives 20, 10 and 60 percent for the first three parts, while UCD gives seconds instead: about 15 for Situation and Task together, 60 to 75 for Action and 15 to 30 for Result. Those ranges sum to UCD's own total of 90 to 120 seconds.1 (right) "Use the split as guidance and keep Action the longest part."
Follow-up questions
How long should the answer to "describe your past projects" be?
UCD recommends a total of 90 seconds to 2 minutes. At MIT's shares that gives 54 to 72 seconds of Action and 18 to 24 seconds of Situation.1
Should I use the STAR method for a project question?
Yes. The Amherst College guidance suggests writing a STAR paragraph and then saying it out loud, which is a useful way to rehearse the answer.
What if my project was a team effort?
Describe the team in one clause, then say what you personally did with "I". A strong answer separates your decisions from the team's work.
What if I have no work projects to describe?
Use any project where you can name your own decisions and a measurable outcome, such as a course or personal project. The same four parts apply.
How do I describe a machine learning or embedding project?
State the metric and its baseline, the data scale, and one trade-off you chose. For background on the building blocks, see similarity search, hybrid search and the two-tower model, which rest on a vector space and the transformer. This advice is our own, drawn from the structure above rather than from a source on machine learning projects.
What if the result of my project was poor?
Report it with the number and say what you learned or changed. A measured poor result with a clear lesson shows more judgment than an unmeasured good one.
- Pick one specific project, chosen to fit the job, not a list of everything you have done.
- Tell it in STAR order, with the Action as the longest part.
- Say "I" for your own steps and add one trade-off sentence.
- End with a number, its baseline and its cost.
- Treat the percentage splits as guidance, because the sources disagree.
Sources
- Using the STAR method for your next behavioral interview — Massachusetts Institute of Technology, Career Advising and Professional Development (accessed 2026-10-26), 26 Oct 2026
- Breaking down the STAR method — University of Iowa Tippie College of Business, Pomerantz Career Center (accessed 2026-10-26), 26 Oct 2026
- STAR Method — Amherst College Career Center (accessed 2026-10-26), 26 Oct 2026
- How we hire — Google Operations Center (accessed 2026-10-26), 26 Oct 2026
- STAR (Situation, task, action, result) — Wikipedia (definition pointer, accessed 2026-10-26), 26 Oct 2026
- STAR Interview Technique — University College Dublin Professional Academy (accessed 2026-10-26), 10 Aug 2026
