MerchantOps

Autonomous payments, dispute resolution, and risk defense for omnichannel merchants.

Comment

Inspiration

Operating an omnichannel business today is an operational nightmare. Merchants sell online through e-commerce and in-person via point-of-sale terminals (PayPal Zettle). Yet, when chargebacks, refund requests, or fraud rings hit, the critical data required to defend the business is scattered across disconnected systems:

  • Carrier tracking portals (FedEx, UPS, DHL)
  • Customer support desks (Zendesk, Gmail)
  • Point-of-sale shift logs (PayPal Zettle)
  • Financial workflows (Invoicing, Mass Payouts, Subscriptions, and 1099-K tax reporting)

Merchants lose hours manually copying tracking numbers, downloading delivery signatures, and cross-referencing order records—often discovering card-testing bots or dispute deadlines when it's already too late.

We built MerchantOps to give merchants a unified, AI-powered operations command center that bridges PayPal’s entire financial ecosystem with automated evidence ingestion, predictive risk defense, and strict human-in-the-loop governance.

What it does

MerchantOps acts as an intelligent operating system for merchants across five core pillars:

  1. Autonomous Dispute Dossier & Evidence Assembly: When a dispute is opened, MerchantOps automatically ingests courier delivery signatures (FedEx/UPS/DHL) and customer email exchanges (Zendesk/Gmail), compiling a complete contestment dossier in under two minutes instead of 30+ minutes of manual work.
  2. Predictive Fraud & Velocity Defense: Real-time scoring monitors rapid IP/device switching, card-testing micro-transactions, and velocity anomalies before settlement batches finalize.
  3. Omnichannel Reconciliation: Seamlessly connects digital e-commerce transactions with physical in-store PayPal Zettle POS registers, synchronizing inventory adjustments and daily settlement ledgers.
  4. End-to-End Financial Operations:
    • Automated draft-to-paid invoicing lifecycles
    • Recurring subscription dunning and churn prevention
    • High-volume batch vendor payouts with cryptographic verification
    • PayPal Working Capital loan tracking and automated 1099-K sales tax liability reconciliation
  5. Human-in-the-Loop Governance & Multi-Tenant RBAC: High-consequence actions (mass disbursements, refunds, dispute submissions) require a dual-button cryptographic SHA-256 confirmation gate and strict role-based access control (Owner, Admin, Finance, Support, Cashier).

How we built it

MerchantOps was built from the ground up as a production-grade enterprise application:

  • Frontend & UX: Built with Next.js 15 (App Router), React 19, TypeScript, and Tailwind CSS. Designed with an enterprise light-mode visual language (benchmarked against Ramp and Wise), featuring a global command palette (Cmd+K), live settlement preview widgets, and domain workspaces (/disputes, /risk, /pos, /payouts, /invoices, /taxes).
  • Database & Storage: Serverless PostgreSQL powered by Neon DB, storing tenant isolation profiles, webhook audit trails, transaction records, and action logs.
  • Multi-Tenant Agentic AI Engine:
    • Python & LangChain orchestration with provider-agnostic LLM support (OpenAI GPT-4o, Anthropic Claude 3.5 Sonnet, Google Gemini 1.5 Pro).
    • Construction-time API key isolation ensuring enterprise data segregation with zero global memory leakage.
    • Safe deterministic calculation tools preventing arithmetic hallucinations for ledger balancing: $$\text{Net Settlement} = \text{Gross GMV} - (\text{Disputes} + \text{Refunds} + \text{Fees})$$
    • Additive lifecycle hooks (pre_hook, post_hook) producing immutable audit trails for every AI action.
  • PayPal & Webhook Integrations:
    • PayPal REST APIs: Invoicing, Orders, Disputes, Payouts, Subscriptions, and Zettle POS.
    • Carrier & CRM Webhooks: FedEx, UPS, DHL tracking milestones and Zendesk/Gmail customer communication threads.

Challenges we ran into

  • Multi-Tenant Key Isolation: Allowing enterprise organizations to bring their own LLM keys (or connect via an OpenRouter gateway) without cross-tenant key leakage or cross-talk required engineering a dynamic factory pattern with zero shared state.
  • Normalizing Unstructured Evidence: Shipping tracking webhooks and email threads follow drastically different schemas. Cleaning, validating, and formatting timestamps, delivery signatures, and customer support messages into a coherent chronological timeline for dispute adjudication was a complex data engineering task.
  • Preventing AI Financial Hallucination: An LLM should never be permitted to execute money movements autonomously. We designed an enforced two-step verification drawer with SHA-256 payload hashing, ensuring the AI assists with synthesis while the merchant maintains absolute authorization control.

Accomplishments that we're proud of

  • Sub-2-Minute Dispute Turnaround: Reduced the time to compile, correlate, and assemble complete dispute evidence packets from 30+ minutes down to under 2 minutes.
  • Provider-Agnostic LLM Architecture: Seamlessly switching between GPT-4o, Claude 3.5 Sonnet, and Gemini models with zero code rewrites.
  • True Omnichannel Integration: Bringing online PayPal transactions and in-person Zettle POS shifts into a single, cohesive operational dashboard with instant reconciliation.

What we learned

  • Merchants prioritize control over pure autonomy: Full autonomy in financial software creates hesitation; transparent, human-in-the-loop approvals backed by explainable AI summaries build lasting trust.
  • Context is everything in risk: A flagged transaction that looks suspicious in isolation often makes total sense when cross-referenced with an in-store Zettle purchase or an active customer support thread.

What's next for MerchantOps

  • Native PayPal MCP Server Integration: Integrating natural-language operations via the Model Context Protocol (MCP) to allow direct action execution from any MCP-compatible agent.
  • Automated Carrier Claim Filing: Automatically filing lost-package reimbursement claims with FedEx/UPS when carrier fault is detected during dispute investigation.
  • Multi-Location Inventory Balancing: Real-time cross-channel inventory rebalancing between online storefronts and physical retail registers.

Built With

Share this project:

Updates

Submission history