Nearhand

Nearhand is scheduling for small businesses, with an AI agent that fills open shifts from the neighborhood. The problem: when someone calls out sick, a café or restaurant owner fixes it by texting around while serving customers, and every unfilled shift costs sales. Big chains have workforce software; a ten-person café has a group chat. How it works: • The owner writes the week's rules the way they'd say them ("@Maya off Tuesdays, a keyholder at every open and close"). Muse Spark Open turns them into typed rules shown as chips the owner can check and untick, and asks a question when a rule is unclear. • OR-Tools CP-SAT builds the cheapest week that follows every rule: quarter-hour shifts, money in integer cents, overtime, rest between shifts, required skills. • A gap-filling agent (Muse Spark Open with tool calling) fills the open shifts. It searches the neighborhood, asks the optimizer, tries what-ifs such as giving a team member extra hours, and skips the ones that cost more. It asks the business's own team first, then qualified part-timers nearby. Every number it writes is checked against the optimizer before the owner sees it, and nothing is booked until each person says yes. • Everyone answers on their phone. When someone declines, the agent offers the shift to the next-best person on its own. "Should I take this?" gives a worker an honest second opinion checked against their calendar, weekly max, pay and distance. • Workers have a voice: a chat message like "can I leave at 3 instead of 5?" becomes a checked, one-click shift change, and the agent re-offers the hours they gave up. • A sick call is one sentence: "Chris called out sick Thursday" reopens his shift and the agent finds cover. The model talks; the optimizer decides. The model never assigns shifts or computes money; code does, and hard limits (calendar clashes, weekly maximums) are enforced in code. Proof: an eval harness checks every model job against known answers. Rules read exactly right 25/25; the agent matched the optimizer's coverage in 10/10 test weeks (within 1% of its cost) at about 2.6¢ per plan; offer advice right 20/20 with no unsafe "take it". 111 automated tests, including full runs for a café and a full-service restaurant. Try it with no sign-up: https://nearhand.streamlit.app ("Take the 2-minute tour", or "See it on their phone" to answer as a worker). All people and businesses in the demo are fictional. Built during the hackathon, Oct 29 – Nov 4, 2026.