Inspiration:
Payment failures and reconciliation mismatches can leave merchants and operations teams asking the same questions: Did the payment actually complete? Is the merchant record accurate? What evidence supports the next action?
ResolveAI was inspired by the need to make payment investigation more evidence-driven, transparent, and safe. Instead of relying on assumptions or immediately changing records, operators should be able to verify payment evidence, understand discrepancies, and make informed decisions.
What it does:
ResolveAI is an AI-assisted payment investigation and reconciliation platform built around PayPal Sandbox.
It helps operators:
- Verify payment evidence: Retrieve and validate PayPal Sandbox payment information before drawing conclusions.
- Detect discrepancies: Compare verified PayPal payment status with merchant-side records and identify mismatches.
- Investigate with AI: Generate structured investigation summaries grounded in available evidence, including missing information and recommended next steps.
- Keep humans in control: Require an authenticated operator and explicit approval before simulated reconciliation actions.
- Maintain accountability: Record investigation and action activity in an audit trail.
The platform is designed to distinguish verified facts from possible explanations. AI should assist an investigation, not invent a root cause or make an unsupported claim about a payment.
How we built it:
ResolveAI combines a JavaScript-based backend, a React frontend, a local database, PayPal Sandbox APIs, and local AI inference.
Technology stack:
- Node.js and Express for backend services and API routes
- React and Vite for the frontend
- PayPal Sandbox REST APIs for payment evidence and verification
- SQLite with better-sqlite3 for application records
- Ollama with Llama 3.2 for local AI-assisted investigation
- AG Grid Community for operational data presentation
- Render for the hosted application
The investigation workflow uses deterministic checks to identify payment discrepancies. AI analysis is performed after payment evidence has been verified. The output is validated, and a deterministic fallback is available when local AI inference is unavailable or its response cannot be safely used.
Reconciliation actions update simulated merchant-side records only. They do not modify a real merchant system or initiate another payment.
Challenges we ran into:
One of the main challenges was ensuring that AI recommendations did not get ahead of the evidence. A completed PayPal payment and a pending merchant record may indicate a reconciliation mismatch, but that alone does not establish why the mismatch happened.
We addressed this by separating evidence verification, deterministic discrepancy detection, AI-assisted investigation, and human-approved action.
Other challenges included handling PayPal Sandbox authentication, validating API responses, maintaining consistent records, preventing duplicate simulated actions, protecting operator access, and making the system useful even when a local AI model is unavailable.
Accomplishments that we're proud of:
We are proud of building an end-to-end payment investigation workflow that combines PayPal Sandbox verification, deterministic discrepancy detection, AI-assisted analysis, and human-approved reconciliation.
Our biggest accomplishment is making evidence and accountability central to the workflow. ResolveAI distinguishes verified payment facts from possible explanations, identifies missing evidence, validates AI responses, and maintains an audit trail.
We also built safeguards around simulated reconciliation actions, including operator authentication and duplicate-action prevention. The result is a practical prototype that demonstrates how AI can support payment operations without taking control away from humans.
What we learned:
We learned that a useful AI-powered financial operations tool needs more than a language model. It needs reliable source evidence, deterministic business rules, clear uncertainty handling, secure access, auditability, and human oversight.
We also learned to treat AI as an investigation assistant rather than an authority: verified facts must remain distinguishable from hypotheses and recommendations.
What's next for ResolveAI:
ResolveAI is a hackathon-stage prototype. Future work includes completing and validating the deployed webhook configuration, improving hosted demonstration-data setup, connecting to real merchant reconciliation systems through authorized integrations, and expanding investigation coverage.
The current implementation uses PayPal Sandbox only. It does not process live payments or automatically change real merchant records.
Built With
- aggrid
- ai
- express.js
- fintech
- javascript
- llama3.2
- node.js
- ollama
- payment
- paypal
- processing
- react
- reconciliation
- render
- sqlite
- vite
