Recall Radar Voice (is anything I own recalled?)

Every year the US government issues over a thousand product recalls: 2.6 million Stanley travel mugs, 1.8 million Ninja pressure cookers, airbag sensors in a 2020 Honda Civic. Almost nobody hears about the one in their own home, because the data sits in three agency databases that expect you to know what to search for. Recall Radar Voice turns that into a conversation. Tell it what you own ("add a Stanley travel mug", "I have a 2020 Honda Civic", or import an Amazon orders CSV), then ask out loud: "is anything in my house recalled?" It checks every item live against CPSC, NHTSA and openFDA, explains each match with its evidence, reads the answer back, and walks you through the remedy: "help me fix one." How AssemblyAI is used: the microphone streams straight from the browser to Universal-Streaming (Universal-3.5 Pro real-time model, 16 kHz PCM, formatted turns). Our server mints a short-lived token behind OAuth, so the API key never reaches the browser and audio never touches our server. Formatted end-of-turn transcripts drive a deterministic spoken-language intent router, no LLM in the loop, that understands "details one" the way people say it. Measured live: the spoken answer starts 0.2 to 1.6 s after you stop talking. The rule it never breaks: it will not say "recalled" unless it can name the CPSC, NHTSA or FDA record it read. Every match carries the recall number, date and matched words; it declines to guess. The same server is a standards-compliant MCP server (Streamable HTTP, OAuth 2.0 + PKCE), so every voice command is a tool call that Alexa+, Claude or any MCP host can make. Live deployment, MIT repo with 25 tests and a deploy smoke, no paid dependencies. Built solo during the hackathon.