A SCALABLE MULTI-SOURCE DATA FABRIC FOR ENDTO-END SUPPLY CHAIN VISIBILITY, INVENTORY INTELLIGENCE AND BUSINESS DECISION ANALYTICS - (SUPPLY SPHERE)

Abstract

A modern supply chain produces data at every step, from purchase orders and supplier shipments to warehouse movements, transport tracking, and retail sales. In most midsized manufacturing and distribution companies, however, this data lives in separate systems owned by different teams. The ERP knows what was ordered, the warehouse management system knows what is on the shelves, the transport provider knows where a truck is, and the sales system knows what customers bought. Because these systems do not share a common view, managers struggle to answer basic questions such as how much stock is actually available across the network, which shipments are likely to arrive late, and which products are about to run out. This paper presents Supply Sphere, a scalable multi-source data fabric that connects these systems and uses the integrated data for inventory intelligence and business decision analytics. The data fabric ingests data through change data capture from the ERP and warehouse databases, scheduled API connectors for supplier portals and the transport management system, and a streaming feed of point-of-sale transactions and GPS pings. All incoming data lands in an Apache Iceberg lakehouse on object storage, where a transformation layer built with dbt cleans, conforms, and joins it. A metadata catalog records the source, owner, schema, and lineage of every dataset, and an entity resolution step matches products, suppliers, and locations that carry different codes in different systems. A semantic layer then exposes consistent business measures such as available stock, days of cover, fill rate, and on-time in-full delivery. On the integrated data, the system runs a set of analytical models. A hierarchical demand forecasting model combines LightGBM with seasonal baselines and reconciles forecasts across product, category, and region. An inventory optimisation module computes safety stock and reorder points for each product and location from forecast error and supplier lead time variability. A gradient boosting model predicts the delay risk of inbound shipments from carrier, route, supplier history, and live GPS progress, and a supplier risk score combines delivery performance, quality rejections, and lead time drift. ABC and XYZ classification groups products by value and demand variability to set review policies.