JAB

Mobile money agents are the small shops across Kenya, and much of Africa, South Asia and Latin America, where people send money from their phone and walk away with cash, or the reverse. To serve everyone, an agent has to hold both physical cash and digital float. When either runs out, customers are turned away and the agent loses income. Most agents decide when to restock by feel. FloatWise is an assistant an agent can talk to in Swahili or English. They can send a voice note, paste raw M-Pesa SMS messages, or photograph a paper ledger. Muse Voice Transcribe handles the audio, and Muse Spark handles the vision and text, turning messy input into structured transactions that the agent reviews and confirms before anything is saved. The key design choice is that the language model never does the numbers. Muse Spark calls our code as tools. A stochastic simulation, fitted to the shop’s daily cash-in and cash-out patterns, paydays and market days, runs thousands of possible days. An exact optimization solver then finds the restocking policy that minimizes expected stockouts. Muse Spark checks each result against the solver’s output and turns it into plain advice: “restock cash by Thursday morning, there’s about a 38% chance you run out before then.” The same pipeline answers what-if questions, like market day or school fees, by changing the simulation inputs and rerunning. We validate the simulator against known closed-form results and test that no plan ever produces a negative balance. FloatWise can also replay an agent’s last month to show what stockouts cost, and match nearby shops with opposite surpluses so they can swap cash and float instead of travelling