PayPulse Copilot

Inspiration

Small teams lose time and money because PayPal disputes, invoices, shipment tracking, and downstream payment commitments are handled in disconnected tools. PayPulse Copilot brings those signals into one calm, auditable workspace without letting an AI make irreversible financial or legal decisions.

What it does

PayPulse Copilot analyzes dispute evidence locally in the browser, identifies likely evidence gaps, and stages a review package for human approval. An AG Grid command center combines disputes, dependent invoices, and staged orders so operators can see deadlines, holds, and next actions in one place.

The demo's AI flags issues such as missing shipment tracking, weak merchant explanations, and transaction mismatches. Its output is advisory: the operator must validate the invoice, carrier, tracking number, and merchant note before any package can be staged.

How we built it

• PayPal Sandbox: verified buyer transaction and dispute state
• AG Grid Community: sortable, filterable operational command center with pinned identifiers and pagination
• Transformers.js: local browser inference using Xenova/mobilebert-uncased-mnli, so case narratives stay on-device during analysis
• Cloudflare Worker / ChatGPT Sites: authenticated server actions, policy enforcement, and deployment
• Node.js tests: approval gates, bounded evidence fields, audit behavior, and no-secret guarantees

PayPal is central to the product. The live demo displays Sandbox transaction 9WU41528W79857207 and buyer-created dispute PP-R-QCQ-10193553, under review for $20 USD.

Challenges we ran into

PayPal's Sandbox API rejected buyer impersonation for the current application with No permissions to set target_client_id. We did not fake the missing permission or simulate a successful API-created dispute. Instead, we created and verified the case through the official buyer Resolution Center, exposed the limitation clearly, and kept the API seed path owner-only.

Local inference also required careful model loading, progress states, and failure handling so reviewers can distinguish a real model result from a static mockup.

Accomplishments that we're proud of

• A real public, interactive prototype
• Real local AI inference rather than scripted labels
• A real AG Grid workflow rather than a static table
• A verified PayPal Sandbox transaction and dispute
• Human approval boundaries for evidence and financial actions
• An auditable architecture that never exposes credentials or tokens

What we learned

Fintech copilots are strongest when they reduce cognitive load while preserving accountable human decisions. Transparent capability boundaries build more trust than pretending every Sandbox permission is available.

What's next for PayPulse Copilot

With the appropriate PayPal buyer-side permissions, the next milestone is approval-gated evidence submission through the Disputes API. We would also add merchant document ingestion, carrier verification, deadline alerts, and richer case-level audit exports.

Sandbox disclosure

This project uses PayPal Sandbox only. No real money moved. No evidence was submitted and no dispute was adjudicated. External evidence submission and financial actions remain blocked in the public demo.

Built With

• ag
• artificial
• cloudflare
• machine
• paypal