How can you design a real-time system to handle failures or downtime of Cognitive Services APIs?