BizFlow AI, business operations with human control

BizFlow AI is an operations layer for small and mid-sized businesses. An owner types an instruction — "process an order for Fatima Noor for 2 units of Product A, generate the invoice, and prepare a confirmation message" — and a team of specialised agents plans and executes it against a real database. Every write is transactional: the order, stock decrement, inventory movement, sale and invoice are committed together or not at all. Afterwards the agents re-read the persisted rows and independently verify them, so the UI only shows "Verified" when the read-back checks pass. Sensitive actions are never autonomous. Customer messages, refunds and restock purchases are prepared into approval requests with the exact payload; a named human approves or rejects them, and the decision is written to an append-only audit log. When a request cannot be completed — for example insufficient stock — the workflow fails safely with zero persisted writes and later steps marked skipped, rather than reporting a partial success. The application is a production-shaped web platform: JWT authentication with bcrypt hashing, tenant-scoped queries, rate limiting, idempotency keys, timeouts and retries, 15 screens covering orders, inventory, customers, finance and reporting, and a FastAPI + PostgreSQL backend deployed with Docker Compose on Vultr Cloud Compute. The AI layer is honest by design. Vultr Serverless Inference is optional for this challenge, so the platform also ships a deterministic planner and states clearly in the UI which mode is active. It never claims a model produced output it did not.