Inspiration
Most AI tools applied to a payment dispute would just write a persuasive sounding response, whether or not the facts actually support it. That's the wrong failure mode. A confident response built on an invented fact is worse than no response at all, since the other party can just check it.
What it does
PayPal Resolve investigates a PayPal dispute and builds a claim by claim evidence table in deterministic server side code, not by asking the AI to decide what's true. NVIDIA Nemotron then drafts the merchant's response as structured claims, each one tagged with the evidence ID it rests on. A server side validator checks every single claim against the evidence ledger before it's allowed into the final response. Unverified, unknown, or missing evidence IDs get rejected, no matter how convincingly the claim is phrased.
How we built it
Real PayPal sandbox integration throughout: OAuth2 client-credentials flow, Orders API (creates a genuine order), Disputes API (a real submission attempt). The dispute scenario itself is a clearly labeled simulation, since PayPal's sandbox gives merchants no API to self-create a real dispute, only a real buyer filing one through checkout can do that.
NVIDIA Nemotron, through Nebius Token Factory, generates the claims as structured JSON instead of free text. That's the key architectural decision: the model's output isn't the final answer, it's a proposal that a separate deterministic check either accepts or rejects.
Challenges we ran into
Getting the model to consistently return valid structured JSON under token pressure took a few iterations. An early version truncated mid-response under a tight token budget, and the validator correctly failed closed, producing nothing, instead of guessing at a partial result.
We also had to be honest with ourselves about a limitation of PayPal's own sandbox: there's no API for a merchant to self-create a dispute, only a real buyer filing one through checkout can do that. Rather than fake it, we labeled the dispute scenario as simulated everywhere, in the UI, the README, and this writeup, while keeping every other API call (order creation, OAuth, the evidence submission attempt) genuinely real.
Accomplishments that we're proud of
Building a real adversarial test for our own system: a buyer message specifically crafted to pressure the model into claiming a delivery happened despite missing evidence. We ran it live, and the validator caught and rejected the unsupported claim, regardless of how the model phrased it. That's not a hypothetical safety claim, it's a demonstrated result against our own live inference.
We're also proud of catching our own mistake honestly. An early version of the claim text parsing wasn't strict enough, and we fixed it before it ever reached a demo, rather than presenting a weaker result as finished.
What we learned
Prompting a language model not to do something is a request, not a guarantee. An independent, deterministic check run after generation is what actually makes a safety property real, not the wording of a prompt.
What's next for PayPal Resolve
Expanding beyond the single dispute reason code tested here, connecting the evidence sources to real systems (shipping APIs, CRM) instead of a demo dataset, and persisting every investigation as a real, query able audit log a merchant could hand to PayPal or a court if a dispute escalated.
Built With
- css
- express.js
- html
- javascript
- nebius-token-factory
- node.js
- nvidia-nemotron
- oauth2
- paypal
- paypal-sdk
- restapi
- stripe-webhooks