Inspiration: When you pay an AI to do work, the AI decides when the job is done. For a small online seller cleaning up a messy supplier catalog, that can mean made-up prices or changed product IDs, and they’ve already paid for it. We wanted payment to depend on proof, not on the AI’s word.
What it does: VoxProof lets you upload a messy CSV file and say what you need out loud, like “Preserve all product IDs. Never invent missing prices.” Google Gemini turns each sentence you say into a rule that code can check. You tap Agree, and PayPal puts the fee on hold without charging it. An AI then does the conversion, and a separate checker written in plain Python, with no AI involved, tests the result against every rule. If everything passes and you tap Accept, the payment is captured. If anything fails, the hold is canceled, and you pay nothing. Voice can explain the results, but it can never move money.
How we built it: The backend is written in Python with Flask, and the frontend is written in plain JavaScript. Google Gemini drafts the contract, plans the work, and answers questions. PayPal handles the payment hold, capture, and cancellation through its Orders and Payments APIs and its checkout button. AG Grid powers the results table, so you can filter failed rows and compare each value with the original file. The browser’s speech features handle voice input and spoken narration. The project includes 51 automated tests and a Postman collection.
Challenges we ran into: The hardest parts were ensuring the AI could never approve its own work or trigger a payment, reading prices written in many different formats without guessing, and keeping the demo running when the free AI model was overloaded. We added automatic backup models to handle that.
What we learned: We learned how PayPal’s hold-then-capture flow can be used to pay only for verified work, and that keeping “AI does the work” separate from “code checks the work” makes AI services much more trustworthy.
