LS Central's POS Stream Is Your Most Valuable Dataset. Here's the ML Layer We Built on It.

Every time a cashier scans an item, LS Central writes a transaction record. In a 10-outlet supermarket processing 15,000 transactions a day, that is 450,000 data points every 24 hours — each one carrying item, quantity, price, time, store, basket ID, and member ID.

Most retailers use this data for reporting. We use it as a real-time feature store for machine learning models that run inside the live operation.

The Pipeline: LS Transaction to ML Inference in Under 200ms

• Event capture: LS Central Data Director pushes POS entries to Azure Event Hub in real time — sub-100ms latency from scan to stream
• Feature engineering: A stream processing layer computes rolling features: basket size so far, time since last scan, item category sequence, current promotion eligibility, store traffic pattern for this time slot
• ML inference: The feature vector hits an Azure ML endpoint running our basket completion model
• Action injection: The inference result feeds back into LS Central's offer engine — the cashier's screen shows a targeted recommendation before the transaction closes

Three Models on This Pipeline

Basket Completion Predictor

Predicts the next 1–3 items most likely to be added based on current basket composition and historical patterns. Precision at top-3: 38%, against a 12% baseline from generic popular items.

Dynamic Offer Trigger

Triggers offers based on basket trajectory rather than time or threshold. Reduced promotional spend by 31% while maintaining the same basket upsell rate.

Shrinkage Anomaly Detection

Monitors cashier and terminal patterns for statistical anomalies — high void rates, items consistently scanned under wrong PLUs, suspicious no-sale patterns. It generates an exception report every morning. Shrinkage identification went from reactive to proactive.