What a structured internship interview can reveal about candidates

we ran an experiment in march 2025. every applicant to our internship program got a 90-second phone screen with a single human — not an algorithm, not a panel, a real person, calling from chennai. the question was the same every time. the answer varied.

the question was: "tell me about a piece of code you wrote that you wish you could rewrite."

14% of applicants got a callback. the other 86% got a polite no. we did not look at resumes, gpa, college name, or branch. we did not look at github. we did not ask leetcode questions.

14% callback is a strange number. it is too high to be random and too low to be a hiring bar. it is the rate at which a 90-second first impression correctly predicts engineering judgment.

what we learned

three things surprised us.

1. the answers that worked had a specific shape. they all had: a problem statement, a decision, a regret, and a reason the regret was load-bearing for the next decision. applicants who talked for 60 seconds or less, with all four parts, converted at 41%. applicants who talked for 90+ seconds, without naming a regret, converted at 2.3%.

1. college name did not predict callback. branch did not predict callback. gender did not predict callback. what predicted callback, in order of effect size, was: specificity of the regret, honesty about the cost of the wrong decision, and a one-line "what i would do differently now" that showed the decision had been revisited. gpa, after controlling for these three, was a small negative (people with a 9+ gpa were slightly less likely to convert, which we attribute to over-preparation and under-reflection).

1. github was almost useless as a signal. we had access to 1,168 of the 1,200 applicants' github profiles. having any github activity at all predicted callback at the noise level. having a personal portfolio site predicted nothing. having open-source contributions predicted callback at the noise level. the only github signal that survived was: a project with a deleted commit, a force-push, and a 3-paragraph readme — the artifacts of someone who had shipped something, then thought it was wrong, then changed their mind.

what we got wrong

we got two things wrong.

first, we did not realize how much the question's framing mattered. the question presupposes that the applicant has a regret. about 18% of applicants did not. we should have had a follow-up: "if you have not written a piece of code you would rewrite, tell me about a piece of code you would not change, and why." we did not, and we likely filtered out thoughtful people who had not yet had the experience the question was probing for.

second, we assumed that the same question would work across roles. it does not. the question works for backend and platform. it does not work for frontend, where the regret is often "i used a css framework that has since been deprecated" — a real regret, but a different shape, and a 90-second answer does not surface the kind of judgment the question is testing for. we had a separate question for frontend and data roles, and we did not write it down.

what we built instead

we replaced the 90-second phone screen with a 20-minute structured interview. the question set is now: three reasoning questions, one short coding block, and one pushback round where the AI disagrees with you and you defend. every answer is scored on a 4-point rubric. every reject has a written reason. every feedback request gets a response within 48 hours.

the 90-second question is still in the interview, as the opener. it is the one piece of the old protocol that we kept.

why we are publishing this

because it is the data we wish we had when we were applicants. because the same dynamic — a 90-second first impression that locks in 14% of outcomes — shows up in every hiring funnel, in every country, in every role. because the question itself is reusable. because we want applicants to be able to prepare for the question instead of being surprised by it.

the question is: "tell me about a piece of code you wrote that you wish you could rewrite."

that is the only question you need to be able to answer in 90 seconds, with all four parts. problem. decision. regret. what you would do now.

if you are applying to dreamclerk, the question is in the interview. if you are applying somewhere else, the question is also probably in the interview, in some form. prepare for it. ship a piece of code you would rewrite. it is the only preparation that matters.

— dreamclerk team, chennai, feb 2026

related posts

• inside our bias audit
• coding interview with no experience