Data science interviews when AI can write the notebook

Define the claim. Separate the evidence. Defend the decision.

A prediction is not an experiment.

Waymark · authored practice data
Change	Human	Simulated
Earlier copy	+2 pp	+1 pp
New fee	Not measured	+3 pp

Open the evidence desk

Practice interviews

01 A useful feature must exist when you decide

“The behavioral half cannot be translated, because no source file tells us how a passenger will behave.” (opens the source)

Henokh Fibrianto and Larry Lin, Co-authors, Grab Tech · Grab Tech, Oct 5, 2026

A later refund can explain a cancellation without helping you predict it.

Data Science · Decision-time dataCan this model predict a cancellation?Example

02 Simulated preference is not a human effect

“Unlike random assignment in a user experiment, neither assumption is guaranteed by design.” (opens the source)

Sebastian Ankargren, Joel Persson and Mårten Schultzberg, Senior Data Scientist; Research Scientist, Computational Economics; Senior Manager, Staff Data Scientist · Spotify Engineering, Aug 13, 2026

A new fee policy needs evidence about people, not just responses from a simulated panel.

Data Science · Human evidenceCan simulated responses justify a new fee?Example

03 A cleaner mean can erase the intended effect

“When the intervention genuinely moves the tail” (opens the source)

Margarida Moreira da Silva, Author, Booking.com ML & DS Blog · Booking.com ML & DS Blog, Jul 1, 2026

Capping legitimate large orders changes what your experiment measures.

Data Science · Define the outcomeDoes capping spend erase the intended effect?Example

04 Your control can share the treatment’s drivers

“seek disconfirming evidence for your hypothesis.” (opens the source)

DoorDash, Global Operating Principles — Truth · DoorDash Careers, Undated

A consumer-level comparison does not remove competition for shared supply.

Data Science · Design the assignmentCan your control share the treatment’s drivers?Example

05 Do not extend a test until it looks favorable

“Did the experiment adhere to its planned runtime?” (opens the source)

Edgar Cano, Daisy Duursma, Nils Skotara and Melanie Mueller, Co-authors, Booking.com ML & DS Blog · Booking.com ML & DS Blog, Mar 24, 2026

Separate an exploratory result from the decision you specified before the test.

Data Science · Keep the decision ruleShould you extend the test until it turns green?Example

Interactive investigationCan simulated responses justify a fee policy?

Sources

Henokh Fibrianto and Larry Lin, Co-authors, Grab Tech, at the time. Grab Tech, “Powering AI-led research through simulation”, Oct 5, 2026. Open the source. Checked against the published text.
Sebastian Ankargren, Joel Persson and Mårten Schultzberg, Senior Data Scientist; Research Scientist, Computational Economics; Senior Manager, Staff Data Scientist, at the time. Spotify Engineering, “When Can LLMs Replace Humans in A/B Tests?”, Aug 13, 2026. Open the source. Checked against the published text.
Margarida Moreira da Silva, Author, Booking.com ML & DS Blog, at the time. Booking.com ML & DS Blog, “Outlier Handling at Scale in Experimentation”, Jul 1, 2026. Open the source. Checked against the published text.
DoorDash, Global Operating Principles — Truth, at the time. DoorDash Careers, “DoorDash Analytics”, Undated. Open the source. Checked against the published text.
Edgar Cano, Daisy Duursma, Nils Skotara and Melanie Mueller, Co-authors, Booking.com ML & DS Blog, at the time. Booking.com ML & DS Blog, “Scaling Experimentation Quality at Booking.com”, Mar 24, 2026. Open the source. Checked against the published text.

Last reviewed Nov 10, 2026