AI-assisted coding interviews for data scientists
More companies now leave an AI assistant open during the coding round and watch how you use it. The notebook below is one of those challenges and you are welcome to work through it yourself.
From remembering syntax to diagnosing a problem
A closed-book coding round tests whether you can produce a groupby from memory. Once the assistant is open that stops being a signal worth collecting, because everyone in the process can get working code out of it in a few seconds. What the interviewer watches instead is your product sense, and specifically whether you know which question to ask next and whether you notice when clean-looking output is answering the wrong question.
That makes these rounds much closer to a root cause case study than to a coding test. The question arrives as a simple evaluation and turns into a diagnosis the moment the data disagrees with the premise, which is where most of the marks are.
Did the initiative improve completion rates?
The product team ran an initiative aimed at getting more people to finish what they start watching, and they want to know whether it worked. That is the entire brief you are handed, and the openness of it is deliberate, because nobody hands you the decomposition in advance.
-
Step 1
Check whether it worked
Have the assistant build a time series and plot it. Establish whether the movement is real or sits inside the range the metric wanders through anyway, and be ready for the answer to disagree with what the product team expected.
-
Step 2
Form hypotheses and segment
A result that contradicts the premise is a starting point rather than a conclusion. Seasonality and a tracking bug leave different fingerprints in the data, and the assistant will test several explanations faster than you can type one of them.
-
Step 3
Follow the hints you are given
Interviewers usually steer. Here you would be pointed toward genre, and that breakdown contradicts the headline number in a way worth recognizing quickly.
Colab runs in the browser with nothing to install. The first cell generates the dataset from a fixed seed, so your numbers will match mine exactly. Nothing is gated and no email address is required.
This is one round out of a loop with six of them
The Playbook covers the AI-assisted coding round in full, along with case studies, statistics, applied machine learning, project deep dives and the behavioral rounds, drawn from the 80+ rounds I went through in Spring 2026.