
Deep learning has not displaced gradient boosting on tabular data, and the candidate who reaches for a transformer on a fraud detection problem is telling the interviewer something about their judgment.
Tabular data has properties that suit trees and frustrate neural networks: heterogeneous feature types, irregular decision boundaries, missing values with meaning, and no spatial or sequential structure to exploit. Trees handle all of that natively. Neural networks need it engineered away first.
Boosted models also train in minutes, tune predictably, and give you feature importance for free, which matters more than a decimal place of AUC when someone has to explain a declined transaction.
Propose boosting as the baseline, say why it fits the data, then name the specific condition that would make you change your mind. That answer reads as experience. "I would use a neural network" reads as a default.
The follow-up is usually about the imbalance, the threshold, or the cost matrix, and that is the real question anyway.

Not keyword stuffing, and not a new resume per job. Reordering the truth, with the numbers attached.

Bigger model or more tokens? The question has a numerical answer, and it comes up in interviews more than people expect.

Five rounds, five different things being graded. Most candidates prepare for two of them and are surprised by the rest.