You have to send a daily email to get reimbursed to your manager if using it. Software managers already have no time in Amazon. Entry level tech people get at least $75/hr straight out of college, so it looks really bad to ask your boss to spend time each day approving the expense to actually receive the perk.

If they start to resent that daily paperwork, you get delayed for promotion or are first on the chopping block during layoffs.

Same thing at amazon. $150/month in rideshare credits but you have to file an expense report for each ride and your boss has to approve it and it just looks bad if you are asking your manager every day for expensing $12 Uber rides.

I think the system has friction on purpose to prevent usage with the illusion of a perk.

Thank you for the clarification.

For what type of problems would this be useful for? Leetcode style problems have deterministic conditions, why use ML to solve that class of problems?

You could iterate through synthetic data experiments until you can produce the output in the correct ratios. This would take a lot of time most likely and could be a false positive

Also multi-instance learning methods could be helpful, in this case to be more resilient to background data.