featurizer

Deep Feature Synthesis compiled to pure PostgreSQL, with point-in-time-correct as-of semantics. 67 aggregations · 83 transformers · leakage-safe by construction.

Start here: the walkthrough View on GitHub Latest release

How it works

One YAML file describes your entities and relationships; featurizer plans, renders, and executes pure PostgreSQL that materializes a point-in-time-correct feature matrix:

Featurizer pipeline: config.yaml → Validator → Planner → SQL Renderer → Executor, with the φ-bridge precompute feeding the Planner and a point-in-time-correct feature matrix as output

Explore

Tutorial notebooks

Five executed notebooks, from basic aggregations to custom primitives — rendered from their validated outputs. Follow the learning path.

References

All 150 primitives with SQL examples (generated from the registry — it cannot drift), the complete configuration schema, and the CLI.

Live-database validation

Every release is validated by materializing full feature matrices against three live PostgreSQL databases. The v1.0.0 matrix is fully green — every variant materializes (including the ~39k-column wide config), and the 0.9.x graph/text families are measured at scale for the first time. Read the reports.

Under the hood

How the SQL stays fast — joins, CTE assembly, set-based pre-aggregation, sharding — plus the φ theory behind feature creation, the ADRs, and the changelog.