ShadowHive lets a person issue commands by silently mouthing words and shows a live indoor map of where help is needed, all over a phone hotspot with no cloud. A single facial EMG electrode on the lips feeds an NXP FRDM-MCXN947, where a causal temporal convolutional network detects that a word is being mouthed and a depthwise-separable CNN classifies it, both running as 8-bit integers in about 15 ms and producing a decision roughly a quarter of a second after the word ends. An iPhone streams its pose and RoomPlan geometry to a Mac, which builds a 2D plan of walls, doors and objects, tracks a UWB-tagged person as a red dot, and plans the shortest route to them. An ESP32 with a 3.5-inch touchscreen renders that plan from a compact scene description, so map updates cost a couple of kilobytes and dot moves a few dozen bytes, and each recognized word appears on it as a banner. Measured on a held-out session, the two-word system gets 88 percent of words right with none missed and one false alarm in seven minutes, and the board reproduces the host's decisions exactly on replayed recordings.