Context
Exact SKU matches were insufficient for identifying viable product alternatives.
Problem
Manual product matching did not scale and produced inconsistent results.
Constraints
Matching needed to balance automation with human review for ambiguous cases.
Scope
AI systems lead responsible for product similarity pipelines.
Strategy
Use embeddings and similarity scoring to automate high-confidence matches and flag edge cases.
Architecture
Built a Node.js pipeline to generate embeddings, update satellites, and compute similarity in link tables.
Impact
Automated the majority of product matching while improving match quality.
Effects
Analysts focused on judgment calls rather than mechanical matching.
Artifacts
Embedding generation pipeline
Available upon request
Similarity scoring model
Available upon request
Key Insights
Human-in-the-loop systems scale judgment instead of replacing it.