Design an end-to-end ML system to detect and take action on ads/listings that attempt to sell weapons (or weapon-related prohibited items).
Your system should cover:
-
Where detection happens (at creation time vs post-publication)
-
What signals/modalities you use (text, images, seller behavior, etc.)
-
How you collect and label training data
-
How you evaluate and monitor the model
-
How you integrate with enforcement (block, downrank, human review)
The interviewer may also ask:
-
How would you collect training data, especially for rare policy-violating content?
-
In a “funnel” setting (impression → click → conversion), is the training data the same for each stage/model? Why or why not?