Write resume project bullets that hold up under follow-up questions

Interviewers usually pick one or two bullets and dig. They are not checking whether the project was impressive. They are checking whether you did it and understood it.

The five-question ladder

For each bullet, the interviewer tends to climb this ladder:

A bullet that invites rung 1 but collapses at rung 3 is worse than no bullet, because it signals you copied the approach.

Rebuild one weak bullet

Starting bullet:

- Built a customer churn prediction model using Random Forest with 92% accuracy.

Run the ladder:

• Rung 1, why Random Forest? "It is good for classification." Weak: no comparison.
• Rung 2, the data? "A Kaggle telecom dataset, about 7,000 rows." Acceptable.
• Rung 3, how did you know? "92% accuracy on test set." Problem: if the churn rate is around 25%, a model predicting "no churn" for everyone gets about 75%, so 92% needs a baseline and a metric that suits imbalance.
• Rung 4, what went wrong? "Nothing really." A red flag. Something always goes wrong.
• Rung 5, differently? "Maybe try XGBoost." Generic.

Rebuilt bullet after actually revisiting the project:

- Built a churn classifier on a public telecom dataset (7,000 customers, ~26%
  churn); compared logistic regression and random forest against a majority-class
  baseline using recall at a fixed 20% contact budget; found tenure and contract
  type drove most of the lift.

Now the ladder has real answers:

• Why: compared two models, chose by recall at a budget.
• Data: size and churn rate stated.
• Measured: named baseline, named metric tied to a business constraint.
• Went wrong: prepare this separately, for example "My first version had a TotalCharges column that encoded tenure, so it looked better than it was until I checked feature importance."
• Differently: "I would model time-to-churn, because the business question is when to contact, not only who."

Note what changed. The bullet is not more impressive. It is more specific, and the specifics are true.

Write the "what went wrong" note for every project

Keep a private file with one paragraph per project covering one real problem you hit. Interviewers ask this often, and the answer shows more skill than the headline result. Examples of genuine issues worth describing:

• A join that duplicated rows and inflated totals
• A date column parsed in two formats
• A feature that leaked the target
• A metric that looked good only because of class imbalance
• A dashboard measure that gave different numbers at different filter levels

Numbers: which ones you may use

The bullet template

[Verb] [what you built] on [data: source, size, key property];
[how you evaluated it: baseline, metric, validation];
[one finding or decision it supported].

It is not mandatory to use every slot in every bullet, but a bullet missing both the data and the evaluation slots will almost always fail rung 2 or 3.

Mock the ladder with a friend

Give a friend the five questions and your resume. Ask them to choose any bullet and read the questions in order, and to push with "why?" if an answer is vague. Time yourself: aim for 30 to 60 seconds per rung. Any bullet where you paused for more than a few seconds needs work, either on the page or in your understanding.

The same ladder applies to a submitted assignment, so pair this with our guide to time-boxing a take-home and the line-by-line mapping in tailoring a resume to a job description. The Data Science programme runs mentor reviews that ask exactly these questions of every capstone before it goes on a resume.