Machine Learning System Design Interview, issue 5, Nov 26, 2025

The Multi-Objective Loss Trap

ML Engineer interview at DoorDash, and the interviewer asks:

โ€œProduct wants to maximize ๐˜œ๐˜ด๐˜ฆ๐˜ณ ๐˜Š๐˜ญ๐˜ช๐˜ค๐˜ฌ๐˜ด. Sales wants to maximize ๐˜๐˜ช๐˜จ๐˜ฉ-๐˜Š๐˜ฐ๐˜ฎ๐˜ฎ๐˜ช๐˜ด๐˜ด๐˜ช๐˜ฐ๐˜ฏ ๐˜–๐˜ณ๐˜ฅ๐˜ฆ๐˜ณ๐˜ด. How do you design the ๐˜“๐˜ฐ๐˜ด๐˜ด ๐˜๐˜ถ๐˜ฏ๐˜ค๐˜ต๐˜ช๐˜ฐ๐˜ฏ to balance these conflicting goals?โ€

A mathematically elegant loss function can quietly destroy your real-time ranking system. Hereโ€™s the architectural fix that actually works in production.

The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.

Read it on Substack

Get the next one

Free on Substack. Unsubscribe in one click.