Conversational fallback with memory; reject zero-amount expenses

Username access, web dashboard and coach chat

Run a local model on ZeroGPU instead of paid inference

Deploy coco-finbot