Tell me about a time you balanced model quality against latency or cost in a shipped product.

This is a senior productionization screen wearing a behavioral hat. Juniors answer it as 'I optimized the model'; the bar is choosing an operating point under a hard constraint and aligning stakeholders to it. The strong answer names the real numbers (QPS, p95 latency budget, cost per request, the quality metric that mattered), shows the tradeoff curve was non-linear, and shows the candidate didn't make the call alone, they framed it so product and finance owned the chosen point. Reserved follow-ups: 'how did you decide the quality bar was good enough' and 'what broke after you shipped the cheaper config,' which test whether they measured the tradeoff in production or just at launch.