Inspiration
Someone says: I paid you. A screenshot is not proof. PaidCheck started from that simple problem: how do you verify a payment claim against real payment records, without trusting a screenshot or a chat message?
What it does
PaidCheck, Nebius edition, verifies payment claims with deterministic rules. A parser extracts amount, currency and payer from free text. The original verdict engine then compares the claim with payment records and returns MATCHED, PARTIAL, NOT FOUND, DUPLICATE or UNCONFIRMED. NVIDIA Nemotron, through Nebius Token Factory, explains the verdict in plain language and suggests values only for fields the parser could not extract. It never decides, and unconfirmed suggestions never change the verdict. Tavily adds public web context about the payer's business, in a separate section marked informational only. Web content can never confirm or deny a private payment. The demo uses test transactions, clearly labelled NOT live PayPal records. The verdict logic is identical to production logic.
How we built it
Python, standard library only for the web demo, with the original deterministic matcher unchanged. Nemotron is served through Nebius Token Factory, model NVIDIA-Nemotron-3-Nano-30B-A3B. Tavily Search provides public context, with a privacy guard that skips search for private individuals. The demo is hosted free on Render. The public repository is https://github.com/ledger-scout1/paidcheck-nebius, commit 2f37e30, with 61/61 tests passing.
Challenges we ran into
The hardest rule was the trust boundary: AI and web search must help, but must never move money decisions. We solved it with a code-level guardrail that discards any model statement contradicting the decided verdict, and with tests proving verdicts are byte-identical with the AI and web layers off. We also fixed a Tavily authentication error caused by using a Nebius key in place of a Tavily key, and we kept the demo honest by labelling all payment fixtures as test data, never as live records.
Accomplishments that we're proud of
A live, public demo that returns a real MATCHED verdict for Paid $180 for the logo design, from John Miller, with a Nemotron explanation and real Tavily context, plus a 61/61 passing test suite on the demonstrated commit.
What we learned
AI is best at understanding and explaining, not at deciding whether money moved. Rules decide. Web search gives context, not evidence. Keeping those roles separate makes the system trustworthy.
What's next for PaidCheck - Nebius edition
Keep the free demo available through the judging period, 1-15 December, and explore a hosted production version on Nebius Serverless Endpoints.
Significant update for this hackathon
PaidCheck was first built on 22 October 2026 for the PayPal AI Hackathon. This Nebius entry is a separate repository, significantly updated in October 2026, inside the submission period: a new Nemotron understanding and explanation layer through Token Factory, a new Tavily public-context layer, a new pipeline combining parser, Nemotron suggestions, the original verdict engine, Nemotron explanation and Tavily context, plus new tests for the trust boundary and a hosted demo.
Links
Demo: https://paidcheck-nebius.onrender.com Repository: https://github.com/ledger-scout1/paidcheck-nebius Video: https://youtube.com/shorts/ulaOGIuUiNg
Log in or sign up for Devpost to join the conversation.