How would you design an auto-scaling architecture for an AI service handling 65k+ transactions per second at Visa?
