📦 ExpenSense — AI-Powered Expense Management Platform

“the agent makes sense of your expenses intelligently”

🚀 Overview

ExpenSense is an intelligent multi-agent reimbursement management platform that automates the entire expense-review lifecycle. The system integrates traditional backend services with LLM-powered agents to deliver accurate, auditable, and policy-aligned expense decisions.

Key capabilities include:

• Automated receipt OCR + validation
• Policy-aware reimbursement decisions
• AI-based anomaly detection (frequency, patterns, merchant reasonableness)
• Admin dashboard for manual review
• Secure file, email, and policy retrieval flows
• Full audit logging for every decision

🧠 System Architecture

ExpenSense uses a layered architecture:

• Presentation Layer — Web UI for employees & admins
• Business Logic Layer — Traditional backend workflow & validation
• Agent Logic Layer — AI agents (Expense, Document, Email, Orchestrator)
• Data Access Layer — CRUD abstraction for Firestore, Cloud Storage, Gmail, Vector DB
• Data Layer — All persistent system storage

Architecture diagram available on page 3 of the final report.

🤖 Agents

• Expense Agent — OCR validation, rule checking (R1–R5), anomaly detection
• Document Agent — Receipt processing, PDF validation, RAG-safe extraction
• Email Agent — Notification delivery, secure outbound messaging
• Orchestrator Agent — Coordinates workflows, routes tasks to specialized agents

📑 Key Features

• Automatic approval for valid requests ≤ $500 (Rule R1)
• Mandatory manual review for > $500 (Rule R3)
• Frequency limit enforcement (Rule R2)
• Strict documentation validation (Rule R4)
• Post-approval financial update + audit logging (Rule R5)
• Policy retrieval using Vector DB with safeguard grounding

🛡 Security & Red Teaming

The system includes a full MAESTRO-aligned Red Team Test Suite:

• 29 total tests across 7 layers and 4 agents
• Tests include hallucination checks, data poisoning, prompt injection, sandbox escape, RBAC enforcement, and inter-agent sanitization
• Final report documents 5 vulnerabilities, primarily RBAC-related, with mitigation recommendations

📁 Project Structure

frontend/       # Web UI (React / Vite)
backend/        # FastAPI server, agents, business logic
services/       # Firestore, Storage, Gmail, Vector DB CRUD interfaces
agents/         # Expense, Document, Email, Orchestrator agents
redteam/        # MAESTRO-aligned test suite

System Overview

📜 Authors

ExpenSense Team (COEN 296 Red Team B): Joshua Lee, Yulin Zeng, Andrew Nguyen, Darren Tang GitHub Repo: https://github.com/andrewqqn/coen296project