
1
1
Looking for experience!

Project Overview 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. Live Demos Dashboard: https://given-laundry-carolina-ancient.trycloudflare.com Admin Panel: https://billing-pictures-authentication-thin.trycloudflare.com 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 the department budget, classifies the expense type, creates a database record, and forwards it to the Policy Checker. Agent 2 — Policy Checker (Llama 3.1 70B via Featherless AI): Checks the request against company policies — including 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): Evaluates and routes the request accordingly: LOW risk (under $500): Auto-approved instantly with zero human involvement. MEDIUM risk ($500–$1,500): Sent to the admin panel for manager review. HIGH risk (over $1,500 or policy violations): Escalated directly 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 secure web interface to approve or reject pending expenses. The panel includes: A Live Feed tab showing all agent activity and decision-making in real-time. A Budget tab featuring department spending bars and quick-action controls. Tech Stack Framework: Band SDK (LangGraphAdapter + InMemorySaver), LangGraph, Python, Flask AI Models: GPT-4o (via AIML API), Llama 3.1 70B (via Featherless AI) Infrastructure: SQLite, Docker Compose, deployed live on a Linux VPS
19 Jun 2026