Almost no labels means you cannot train a classifier and you cannot honestly report recall. The senior answer is to score by typicality, threshold by the customer's review budget, and build the labeling flywheel that turns this into a supervised problem within two quarters.
Updated Aug 2026 · Grounded in real Forward Deployed Engineer interview loops and written to a senior-engineer editorial bar.
Almost no labels means you cannot train a classifier and you cannot honestly report recall. The senior answer is to score by typicality, threshold by the customer's review budget, and build the labeling flywheel that turns this into a supervised problem within two quarters.
The trap here is candidates who jump to isolation forest or an autoencoder and stop. The thing interviewers at Stripe and Datadog are actually screening for is whether you understand that without labels you cannot estimate recall, so you must anchor the threshold to a review budget (precision-at-k) and treat every analyst decision as a label you are harvesting. The flywheel beats the model choice.
No comments yet — be the first to share your approach.
