Full on-board model with access to all low-latency data streams. May or may not include "live" training (training might be batched to a server instead, with weights updated every hour or every day, etc.).