Author: Duncan Gilchrist and Jeremy Harman
The first edition DelfinaMay 29, 2024.
When I (Duncan) was on the Marketplace team at Uber, we used to (half) joke that we were facing one crisis after another.
Our core machine learning products directly managed billions of dollars of company funds: targeted promotions, surge pricing, driver incentives, ETAs, pool matching, prepaid rider fees, subscription upsells – the list goes on. We were worried they were fundamentally broken and would sink our business.
Every week a new potential data science disaster emerged. The chaos always started the same way: someone found something suspicious in the data. Maybe it was a spike in an internal metric, like “zeros,” which measures riders who opened the app and couldn't find an available vehicle. Were we missing out and losing riders because of unreliability? Had we rolled out too many rider promotions? Were our driver incentives wrong?
Or maybe it was a tweet, maybe a celebrity thought the fare was too high and walked 20 feet to see the fare change drastically, and then, of course, the screenshot went viral and started going viral.
To continue reading this article, click here.
