Superstore
Superstore replaces in-store surveillance with simulative agents. We built a synthetic shopper population that A/B tests stores before touching a shelf.
Walmart spent $7B optimizing store layouts using customer tracking. But it's slow, expensive, and increasingly illegal as privacy laws tighten.
Inspired by Stanford's Generative Agents paper, we built a synthetic shopper population: each AI agent has a persona, budget, and shopping list, and navigates a real store grid using a utility function derived from the Larson et al. (2005) consumer corpus from Wharton.
Retail surveillance is becoming legally radioactive. Half of US states now have biometric privacy laws, and the EU's AI Act categorizes in-store tracking as 'high risk.' Yet store layout optimization is a multi-billion-dollar industry that depends on knowing where shoppers go.
Each agent has a persona (demographics, budget, dietary preferences), a shopping list (generated from a marketing taxonomy), and a navigation policy (utility-maximizing path through the store grid).
Drop in a product catalog and store blueprint, and in 90 seconds you get heatmaps, conversion metrics, and ranked interventions — move this aisle, reprice this SKU — with projected daily lift.
Working demo on a real CVS-style store layout. Two heatmap visualizations: aggregate foot traffic, and intervention lift. Ranked list of layout changes by projected daily revenue.
Scaling agent population beyond 10k shoppers per simulation. Pitching to a real retailer for a paid pilot.
