SKU Matching Across Fragmented Retail Data
A fine-tuned language model and entity-resolution system matching fragmented product records at scale.
Technology
RPython
Selected project note
The business context
A retail analytics workflow required aligning product-level transaction data from informal vendors with a central product catalogue. Source data was highly unstructured, with inconsistent naming conventions, missing identifiers, and significant variation in how products were recorded across vendors. This made reliable aggregation and analysis difficult, particularly when integrating data across multiple sources.
What was delivered
Developed an NLP-driven matching system that transforms raw product descriptions into structured attributes (e.g., brand, variant, pack size, weight) using a fine-tuned language model. These structured representations were then used within a scoring-based entity resolution framework to match products against a master catalogue, with confidence scoring and fallback logic to handle incomplete or ambiguous cases. This shifted the problem from fragile string matching to a more robust, attribute-based approach.
The practical impact
Improved matching accuracy and consistency across fragmented data sources, while eliminating reliance on brittle rule-based approaches. The solution enabled scalable integration of new vendor data and more reliable downstream aggregation and reporting. It also reduced manual intervention required to maintain matching logic as new product variations emerged.