A Causal Inference Toolkit for Google’s Data Science Team
A Cornell MPS sponsored capstone for Google’s Data Science team: a modular Python toolkit for causal effect estimation, referred to as "BETA" in the anonymized project brief.
The problem
Teams doing causal analysis inside Google's Data Science org repeatedly rebuild the same estimation methods - matching, regression, instrumental variables, difference-in-differences - in ad hoc ways, without a shared, well-tested package or a way to validate methods against known ground truth.
What the team is building
- A modular Python package covering Matching, Regression, Instrumental Variables, and Difference-in-Differences estimators behind a consistent interface.
- A simulation generator, so methods can be validated against data with known, ground-truth causal effects before being trusted on real data.
- Agentic documentation generation, so the package's docs stay in sync with the code as the toolkit evolves.
Key decisions
- Consistent interface across methods - standardized the API across very different statistical methods so a data scientist can swap estimators without relearning a new interface each time.
- Simulation-first validation - built the ground-truth simulation generator early, so the package's correctness could be checked against known effects rather than only against real-world data with no ground truth.
Outcome
This capstone is in progress for Fall 2026. Final results, benchmarks, and adoption details will be added here once the engagement wraps - get in touch for the current status.