how do you handle latency and bandwidth variability across WAN peers to keep inference smooth for real-time apps? | discoverkit | discoverkit