Design the Output Layer and Loss for Three Tasks

For each task, specify the number of output units, the final activation (if any), the loss function, and what to apply at inference. Name one mistake you would flag in code review.

1. Predict delivery time in minutes. Values are positive and right-skewed (most under 40, a long tail to 180).
2. Tag a news article with any subset of 5 topics (an article can have several tags or none).
3. Classify a support message into exactly one of 10 intents.
4. (Bonus) Detect fraud where only 0.5% of examples are positive: how would you initialize the output bias?