The API/online model inference definitely runs through some kind of edge safeguarding models which could do this.