DisputeAI

Inspiration Every month, finance teams at companies like Walmart, PayPal, or even small startups waste hundreds of hours manually cross-referencing invoices against budgets. It is a soul-crushing, repetitive task where a single missed decimal point can cost a company thousands of dollars. We realized that while "sending" money is easy, verifying bills is hard. We wanted to build a "Vibe Coded" solution that acts as an tireless auditor—an AI agent that lives in the email inbox, reviews every line item, and stops unauthorized spending before it happens. What it does DisputeAI is an autonomous accounts payable agent.

1. Watches & Extracts: It monitors an inbox for incoming invoices (PDFs or text). Using AI, it extracts structured data: Vendor Name, Total Amount, Date, and Line Items.
2. Enforces Policy: It runs the extracted data against a logic gate (e.g., "Is this invoice over $500?").
3. Takes Action: ◦ Safe Invoices: Are automatically logged as "Approved." ◦ High-Risk Invoices: The agent automatically flags the transaction, calculates the variance, and drafts a dispute email to the vendor asking for clarification.
4. Visualizes: All data is pushed via API to a secure Mocha dashboard where human managers can review the "Disputed" queue and approve the AI's draft with one click. How we built it We utilized two powerful Sponsor Tools to build a "Pro-Code" architecture without the boilerplate:
5. Relay.app (The Brain): We built the core logic agent in Relay. It handles the Gmail trigger, the AI entity extraction, and the conditional logic (If > $500). We used Relay's HTTP Request feature to POST this structured data out to our app.
6. Mocha (The Interface): We used Mocha to generate a full-stack Fintech dashboard. We prompted Mocha to create a database schema for Invoices and an API endpoint to receive the data from Relay. This allowed us to have a production-grade UI with login, search, and status tracking in minutes.
7. REST API: The "magic" happens in the connection. We engineered a seamless JSON handshake between Relay's automation engine and Mocha's database. Challenges we ran into The biggest challenge was data structure interoperability. Getting the AI extraction from the email (which is unstructured text) to match the strict JSON schema required by the Mocha database API took several iterations. We had to fine-tune the Relay AI prompt to ensure it always outputted numbers as integers (e.g., 500) rather than strings (e.g., "$500") to prevent API errors. Accomplishments that we're proud of We are proud that we built a functional, full-stack Fintech tool with a working API pipeline in under 24 hours. We didn't just build a frontend; we built a working data pipeline that solves a real-world enterprise problem (Spend Management). We also successfully used Mocha's platform credits to deploy a live environment. What we learned We learned the power of Composable AI. Instead of writing a Python script to parse PDFs (which is fragile), we used Relay's pre-built AI nodes. Instead of spending days on React components, we used Mocha. This hackathon taught us that "Vibe Coding" isn't about not coding—it's about architecting systems at a higher level of abstraction. What's next for DisputeAI • Historical Memory: Integrating Pinecone to compare new invoices against past prices to detect inflation trends. • Payments: integrating Stripe to auto-pay the "Approved" invoices. • Multi-Channel: Adding Slack/Teams notifications for the "Approval" step.

Built With

• api
• automation
• fintech
• mocha
• openai
• relay.app