MIRROR Voice Agent Safety Layer

Voice agents jump from language to action. Ask one to "move 700 units of XR-17 from Mumbai to the Pune site" when there are two Pune sites, and it guesses, then writes to your inventory system. A wrong guess is only found when the stock is in the wrong place. MIRROR puts a shadow world between the voice agent and the system of record. Every reading of a command is simulated on a frozen copy of the ERP: Pune Site 1 would go 400 over capacity, Site 2 is fine, so MIRROR asks only the question that changes the outcome. Operators explore by voice ("what if I move 900?", "compare 500 and 900") and nothing real moves. Only an explicit "commit" executes, bound by a signed token to the exact version they saw. MIRROR then reads the whole ERP back: if it short-ships, the screen says "Reality diverged" and a rollback restores and re-verifies. If the ERP changes under an open preview, the agent says so unprompted. Built on AssemblyAI's Voice Agent API: recognition, turn-taking, barge-in, tool calls and speech over one server-side WebSocket. Tool calls are only proposals: held until the reply finishes, dropped if the user interrupts, and every value is checked against the user's own words. Free-form typing goes through the LLM Gateway behind the same checks. Proof: 116 tests, live runs on AssemblyAI (53/53 and 28/28 checks), and on 26 scenarios a naive voice agent took 9 unsafe actions while MIRROR took none. Who it is for: operations teams where hands are busy and mistakes are costly, starting with warehouses. AI agents are forecast to grow from $7.8B in 2025 to $52.6B in 2030 (MarketsandMarkets). Revenue from operator seats, verified transactions and ERP connectors. Try it: watch a replay of a live session, talk to it, or type. Every visitor gets a private sandbox.