Three projects that take messy operational data and reduce it to something a team can act on. The churn model is built to be interrogated rather than admired — cross-validated, imbalance-aware, and explained per customer with SHAP so the output survives a conversation with the people who have to use it. The retail analysis works in the other direction, segmenting customers and cohorts before modelling where discounting stops paying for itself.
The third closes the loop on getting the data at all: a live WebSocket ingestion service that batches, deduplicates and lands trades in PostgreSQL, containerised so it runs the same way anywhere. Together they cover the full path — capture the data, understand it, predict from it.
File
03 entries