What considerations are necessary when deploying a model that must maintain low latency on an endpoint agent at SentinelOne?