Top-10 Latin American retailer

The retailer's inventory forecasting system powered purchasing decisions for over 4,000 SKUs across 200+ store locations. Three months after the ML team deployed a new demand forecasting model, forecast accuracy began a slow, unexplained decline — from 85% to 71% over six weeks. Replenishment errors were costing an estimated $800K per quarter in overstock and stockouts. The root cause was distributed across three data layers: a change in the ERP system's date formatting corrupted the Bronze layer, a downstream aggregation bug in the Silver layer amplified the error, and the Gold layer's feature engineering silently compensated — producing plausible but systematically wrong forecast inputs. No single team saw the full picture.

NEXUS™ was configured with the full three-tier Medallion architecture. Statistical drift detectors were deployed on 47 inventory and sales features, establishing baselines from 90 days of historical runs. The ML Reliability Score (MRS) was computed daily against each feature group. Within the first week of deployment, the temporal drift comparison identified z-score shifts above 3.2 standard deviations in three date-derived features — precisely the ERP formatting change. SENTINEL™ traced the data lineage graph to identify all downstream datasets and models consuming the corrupted features, automatically triggering pipeline isolation and notifying the data engineering team. The Medallion validation thresholds were tuned: Bronze at 50% (flexible for raw ingestion), Silver at 75% (strict on cleaned features), Gold at 90% (non-negotiable for model inputs). Any Gold run failing the 90% threshold automatically blocked the forecasting model from consuming updated features.

Full deployment with Iceberg integration completed in 16 days. The time-travel capability proved immediately valuable — the team used Iceberg snapshot history to reconstruct exactly which pipeline run first introduced the corrupted features. The Apache Iceberg merge mode in the Silver pipeline enabled automated data repair: the team corrected the ERP date formatting, replayed affected runs, and merged corrections without full dataset reprocessing.