AI Expense Approval Pipeline on Band

Enterprise expense approval is slow, error-prone, and creates manager bottlenecks. Our system replaces the manual process with a 4-agent AI pipeline running on the Band platform, combining two AI providers in a single workflow. HOW IT WORKS: When an employee submits an expense via the web dashboard, the request enters the Band room and triggers the pipeline automatically. Agent 1 — Budget Checker (GPT-4o via AIML API): Validates department budget, classifies the expense type, creates a database record, and forwards to Policy Checker. Agent 2 — Policy Checker (Llama 3.ecks the request against companypolicies — travel advance notice, IT pre-approval for software, asset tracking for hardware, and CFO sign-off for amounts over $5,000. Agent 3 — Risk Evaluator (GPT-4o):dingly. LOW risk (under $500,to-approved instantly with zero human involvement. MEDIUM risk ($500–$1,500) goes to the admin panel for manager review. HIGH risk (over $1,500 or policy violation) is escalated to the CFO. Agent 4 — Approval Notifier (GPT-4o): Finalizes all decisions, updates the audit trail, and posts a formatted notification back to the Band room. ADMIN PANEL: Managers access a web interface (npprove or reject pending expenses.The panel includes a Live Feed tab showing all agent activity in real-time, and a Budget tab with department spending bars and one-c TECH STACK: Band SDK (LangGraphAdapter + InMemorySaver), GPT-4o via AIML API, Llama 3.1 70B via Featherless AI, LangGraph, Python, Flask, SQLite, Docker Compose, deployed live on a Linux VPS. The system demonstrates that multian handle real enterprise workflowswith production-grade reliability.