Team Prep Lab

Restaurant Prep Lab helps a small restaurant decide what to prep tomorrow from its sales history, the ingredients it actually has, and its kitchen's limits, so it wastes less food without running out. PROBLEM: Independent restaurants throw away prepared food every night, yet still sell out of popular dishes. Owners guess prep quantities, and stock is hard to track. An invoice shows what was bought, not what's left on the shelf, and small items like cheese and spices quietly run out. WHAT IT DOES: 1. Explore: a year of (fictional) pizza shop sales: best sellers, revenue, and rush hours by weekday and hour. 2. Meta Evidence: the owner writes a quick stock count in plain words ("Dough: ten kilograms on hand"). Muse Spark 1.3 turns it into structured amounts, with an exact quote from the note as evidence for each number. The owner reviews and corrects every amount before confirming. Nothing reaches the planner unconfirmed. 3. Prep Planner: deterministic Python forecasts demand and checks it against fixed recipes and the confirmed stock. Example: 29 pizzas expected, only 20 feasible because of a 900 g cheese shortage, and all 29 feasible after replenishment. HOW WE USED META AI: Muse Spark handles the messy, human part: turning a free-text stock note into a strict JSON schema. We use structured outputs with schema validation, quote matching against the source text, and a finish/refusal check so a cut-off answer is never trusted. During testing we found Muse Spark uses about 1,000 reasoning tokens before answering, so we sized the output budget for that. All business math stays in code; the model never invents a number. TECH: Python, Streamlit, SQLite, pandas, Altair, Meta Model API (Muse Spark 1.3, structured outputs). 97 automated tests. HONEST LIMITS: All data is fictional (a public demo pizza dataset), and recipes and stock are demo assumptions. No real restaurant has piloted it yet, so we make no savings claims. Next step: a pilot with a real restaurant. TEAM: Arshnoor Bhatia (data, forecasting, planning) and Aishwarya Srivastava (Meta integration, review flow).