How would you handle the trade-offs between model latency and output quality in a real-time production environment at Google DeepMind?
