A voice-first double-entry ledger that fact-checks your spending against the live web. You say "grabbed the Dyson V15 Detect for nine forty nine ninety nine." Speechmatics transcribes it as you speak, the receipt prints word by word. Claude turns it into a balanced double-entry transaction. Then Bright Data queries Google Shopping live and tells you the market median is $794.99: you paid $155 over. Say three purchases in one breath and it books three separate balanced transactions. Ask "should I buy the Dyson V15?" and it books nothingm it just tells you what a good price would be. The hard part isn't the voice. The raw Google Shopping median for AirPods Pro is $103.75; they cost $199–249. A pipeline that trusted that would report a 127% overpay. So code decides which listings are even the same product, the model judges only condition, and SQL computes the verdict, and when the evidence is too thin, it prints INSUFFICIENT EVIDENCE and no verdict at all. That refusal is the feature. Underneath, the books can't be wrong: the balance rule is a constraint trigger in Postgres, not a check in our code. Open the URL and you have an account — no signup.
Category tags: