Retail Execution AI - Shelf Recognition & Display Scoring
Shelf and cooler photos in, a display-compliance score out: SKU recognition, a deterministic rule engine, and human review only where doubt could change the payout
Overview
A brand that pays stores to display its products has to check, store by store, whether the display was actually there - and today that check is a person with a clipboard. Retail Execution AI turns it into photos. Photos of the shelf, cooler or counter are submitted through the API; the server checks resolution, sharpness, brightness and glare, rejects duplicates, and blurs faces before anything else touches the image.
Products are detected, grouped into shelf rows with facing order and gaps, and matched to SKUs by embedding search against an index each brand builds from its own packshots - adding a SKU means uploading images, not retraining a model.
What was seen becomes a set of Visit Facts; what it means is decided separately by a versioned, deterministic rule engine with KPIs for presence, facings, share of shelf, position, cooler purity, POSM and price, plus gates, exclusions and store tiers.
Because recognition is never certain, the engine computes the probability that each KPI really passed and sends a visit to a reviewer only when the doubt could flip pass or fail; the reviewer's labels re-score the visit and enrich the SKU index.
Fraud signals - spoofed GPS, outside the geofence, impossible travel speed, reused photos - feed a risk gate, and reward periods run from cutoff through multi-level approval to an Excel export. Each brand is a tenant isolated by PostgreSQL row-level security.
The pipeline runs end to end today on interim open models, which will be replaced by in-house models before commercial use; the same core is planned to be sold as an API for DMS/SFA software, as a SaaS for trade-marketing teams, and as a managed audit service.