OpEx Agent

The autonomous finance copilot that stops reimbursement leaks and compliance violations before they happen.

Comment

Testing Credentials & Test Files for Judges

To experience the autonomous workflow and the Finance Command Center, please use the following demo accounts at our provided app link.

** How to get the Test Receipts:** We have provided sample receipts in the Image Gallery at the top of this Devpost page. Please download them (Right-click -> Save Image As) to test the app:

  • Use receipt-holiday.png to trigger the Custom API compliance violation (The AI will flag it as Christmas Day).
  • Use receipt-pass.png to test a successful, normal reimbursement extraction and approval.

1. Employee Account (To test AI Audit & Upload Receipts):

  • Email: jane.smith@company.com
  • Password: janesmith123
  • Action: Upload the downloaded receipts here to see the AI autonomously process or flag them.

2. Admin Account (To test HitL, Override, & Stripe Payout):

  • Email: admin@company.com
  • Password: admin123
  • Action: Log in here to review flagged claims, read the JSON Audit Trail, and trigger the Stripe payout simulation.

Inspiration

Corporate reimbursement is fundamentally broken. It usually relies on manual data entry by employees, slow approval bottlenecks by managers, and reactive (often too late) compliance checks by the finance team. Fraud and policy violations, such as claiming expenses on public holidays, often slip through the cracks. We built OpEx Agent because we wanted to shift operational expense management from a reactive, manual process to a proactive, autonomous ecosystem powered by AI.

What it does

OpEx Agent is an Autonomous Operational Expense & Audit Copilot. It eliminates data entry and automates the audit process. When an employee uploads a receipt image, our 5-step AI workflow takes over:

  1. Validates the receipt image.

  2. Extracts and sanitizes data (Merchant, Amount, Date) using PaddleOCR.

  3. Audits the amount against the employee's remaining budget.

  4. Compliance Check: Cross-references the exact date with a Custom API to detect policy violations (e.g., claiming expenses on national holidays).

  5. Approves the claim automatically or flags it for manual review.

For the flagged claims, we built a Finance Command Center, a Human-in-the-Loop (HitL) dashboard where admins can review the AI's exact JSON audit trail, make override decisions, and trigger a Stripe payout simulation.

How we built it

We built the entire application using MeDo, leveraging its powerful integration with Baidu's LLM, Supabase, and React.

Structuring the Conversation with MeDo: Instead of blindly prompting MeDo to "build an expense app," we used a highly structured, architectural approach. We started by feeding MeDo a strict prompt to generate a comprehensive Product Requirements Document (PRD). Once MeDo established the business logic, user roles, and the 5-step autonomous audit workflow in the PRD, we executed a "Code Generation" phase. Later, we used precise delta-updates (e.g., "Update Step B to strictly format the OCR date to YYYY-MM-DD") to refine the application without wasting credits on complete regenerations.

API Integration & Plugins: To extend the AI's functionality into real-world corporate compliance, we integrated a Custom PHP API (kalnas-224014.appspot.com). During the AI Audit step, MeDo passes the sanitized date to this API. If the API returns that the date is a public holiday (e.g., Christmas Day), the AI autonomously stops the approval process and changes the status to 'Flagged', leaving a detailed reasoning log for the Admin.

Challenges we ran into

Our biggest challenge was ensuring strict data formatting between the AI's OCR extraction and our Custom Compliance API. Real-world receipts have messy date formats (e.g., "25 Des 26" or "25/12/2026"), but our API required strict YYYY-MM-DD inputs. We solved this by instructing MeDo's LLM to act as a data sanitizer, formatting the date perfectly before making the external API call. Additionally, we ran into production build issues regarding static asset paths (missing app icons), which we efficiently resolved by prompting MeDo to convert static image tags into inline SVG components.

Accomplishments that we're proud of

We are incredibly proud of the Audit Trail Transparency feature that MeDo helped us generate. Instead of a "black box" AI that just says Approved or Rejected, MeDo perfectly generated a UI that exposes the chronological decision-making process—including the raw JSON payloads—right next to the receipt image. This provides absolute transparency for corporate auditors. We are also proud of successfully implementing a true Human-in-the-Loop architecture that balances AI autonomy with human oversight.

What we learned

We learned that building with an AI agent like MeDo requires the mindset of a Software Architect, not just a coder. The quality of the generated application is directly proportional to the clarity of the initial business rules and constraints provided. We also learned how seamlessly Supabase Edge Functions can handle complex, multi-step AI logic securely on the backend.

What's next for OpEx Agent

The next evolutionary step for OpEx Agent is direct integration with enterprise ERP systems (like SAP or Oracle) for seamless ledger syncing, and moving the Stripe payout simulation into live production API calls to fully close the loop on employee reimbursement.

Built With

Share this project:

Updates