Jerry Okafor

Kadent is a family app for shopping lists, receipts and a budget, over one household's data. In this hackathon I added Meta's Muse Spark to it. Every AI feature below was built during the event, in the public repo kadent-ai (first commit Oct 27); the app it plugs into existed before. Scan a receipt and Muse Spark reads the photo or PDF into structured lines. When the lines add up to the printed total, the receipt files itself, with eight seconds to undo, and one write updates the list's prices, the household's purchase history and every budget envelope it touched. If the model can't read it, the phone's own reader takes over. From that history, an agent plans next week's trip. It calls four tools over the household's own data (buying rhythm, store prices and what's left in the budget) and gives a reason for every row; the family can change the draft in their own words, and nothing is written until they add it. A family can also say what they need in one sentence and have it land on the right lists, type a spending note any way they'd say it, ask questions about their spending (a second agent, with seven read-only tools, where every amount on screen comes from Kadent's data, never from the model), and get one written card a week about what changed. Scored on 45 real Canadian receipts against the paper, Muse Spark got 223 of 226 lines and all 45 totals right, against 210 of 226 lines for the phone's own reader. The Meta key never ships in the app: every call goes through a Firebase Cloud Function that checks the person is signed in and applies a quota. AI is off until the family turns it on, and each call sends only what that feature needs.