Feature stores address the critical issue of feature inconsistency between ML model training and serving. Feast is an open-source solution designed to bridge this gap by managing feature registries, materialization, and online retrieval. This ensures models consistently access the same data in production as they did during training. While often demonstrated with relational databases or cloud data warehouses, this article explores an alternative approach. It utilizes Neo4j Desktop as the engine for feature computation. Five of the features involve straightforward aggregations that are easily handled by Cypher queries. The sixth feature, a customer community ID, is derived using graph algorithms. These algorithms identify customers with similar purchase patterns through Graph Data Science node similarity. Subsequently, Louvain community detection groups these similar customers. Feast then ingests all six features. It materializes them into a SQLite online store for efficient serving during inference. This process also ensures the construction of point-in-time correct training data.