Every Won Heard - Voice Marketing Consultant

When a shop owner says her marketing budget out loud, the number should come from her words, not from a model. Every Won Heard is a voice marketing consultant built on that rule. The agent asks only for what's missing (the shop, the neighborhood, the main problem, the monthly budget) and returns a priced 30-day plan from MarketPilot's real service catalog, ending in a checkout link. No tool takes a number. AssemblyAI already infers tool arguments only from user turns and tool results (Aug 26); for money we go further. The model passes the owner's words, and our server reads the amount from the transcript it received. A range or two amounts go back as an error, and the agent asks for one number. Every amount lands in a money ledger with its evidence: the owner's words, when it was heard, when the agent read it back, and when the owner said yes. Only that yes makes it the budget. Prices and totals come from the catalog, never from the model. After an English call, each confirmed amount is matched against AssemblyAI's own session record. English calls run on one stored Voice Agent API agent that moves through five stages. Each stage reveals only the tools it needs and changes what the agent listens for (key terms, transcription prompt, max accuracy while money is said) with session.update. Barge-in cuts the agent off about a second after the owner starts talking. The Voice Agent API has no Korean, so Korean calls run on Universal-3.6 Pro streaming, with the agent's next question sent as agent_context mid-call, and land in the same ledger. In our recorded English call (103 s), a range was rejected, 480,000 and then 380,000 won were read back and confirmed, and the receipt showed 2 confirmed, 2 matched, 1 tool call rejected. We start with cafes, restaurants and salons, and the agencies that serve them. In our video the callers, the Korean consultant voice and the narration are synthesized. The agent, transcription, tools and prices are live.