Inspiration
For a freelancer, “approved” and “paid” can sound dangerously similar. Starting work on uncertain evidence can create avoidable disputes. PayProof makes the evidence state visible before offering advice, so a confident sentence cannot silently become financial authority.
What it does
PayProof is a local application for reviewing payment evidence. Its interactive synthetic scenarios distinguish buyer approval, capture status, pending evidence, conflict, and unknown outcomes. A separate recorded panel shows a real historical PayPal Sandbox order read without making a fresh provider request.
The recorded read returned CONFLICT. The local model was ready, but advisory priority remained unknown and no payment was verified. That uncertainty is shown in the product rather than replaced by a success claim. PayProof cannot initiate a payment, capture, or refund from the demonstrated interface.
How we built it
The read-only service integrates the PayPal Orders API with evidence validation and a SQLite ledger. PayPal order and capture semantics are central to the problem the application addresses. An optional local multinomial Naive Bayes assistant uses a pinned, bundled synthetic corpus. It is a learned model built with the Python standard library, not an external AI service. Source validation precedes advice, and the assistant can abstain when evidence is invalid or unconfirmed. The assistant has no financial tools.
The frontend uses HTML, CSS, and JavaScript. Python serves the local UI. The default fixture uses deterministic wording; the learned assistant is an explicit opt-in. The demonstration separates synthetic behavior from the historical real Sandbox result so viewers can assess each on its own evidence.
Challenges and what we learned
The hard problem was maintaining the boundary between a useful explanation and a verified payment fact. A ready model is not evidence of a successful transaction. Conflicting provider facts must remain visible, and the original outcome must survive later presentation improvements. We also learned to make recorded proof legible on narrow screens without blending it into an interactive simulation.
Accomplishments
We built an interactive evidence workflow, a local advisory model, a read-only PayPal integration, a sanitized recorded-result panel, and a short English film showing the application functioning. Existing recorded-panel checks covered 390×844 and 1440×900. The film lasts 2:40. These are scoped results, not a claim that the prototype is production ready or that every financial path has passed.
What's next
Improve the explanation of conflicting evidence and broaden deployment and accessibility validation. Any further real Sandbox experiment needs separate admission and secure custody. Successful capture, authentic webhook delivery, and a production payment lifecycle are not demonstrated by this entry material.
Log in or sign up for Devpost to join the conversation.