How Didi and Ant Financial Co‑Built SQLFlow to Bring AI to Data Analysts

Background

Oracle highlighted in 2018 that while many developers build AI services with Python or C++, most business analysts work primarily with SQL. In July 2019 Ant Financial open‑sourced SQLFlow, a framework that translates SQL programs into Python code, invokes database and AI engines, and delivers end‑to‑end AI capabilities directly from SQL.

Collaboration between Didi and Ant Financial

The joint effort follows a three‑step plan:

Didi contributes deep domain knowledge of its transportation and finance products.

Didi contributes three high‑value models—a DNN classification model, an explainable model, and an unsupervised clustering model—to SQLFlow.

Didi joins the broader SQLFlow open‑source community to co‑build the ecosystem, sharing models, culture, and practices.

Contributed Models and Use Cases

DNN classification model : powers product‑growth recommendation and targeted user recommendations.

Explainable model : provides SHAP‑based visual explanations for predictions, improving interpretability for operations and marketing.

Unsupervised clustering model : analyses driver activity patterns to generate dispatch suggestions and balance supply‑demand.

Technical Enhancements Added by Didi

Integrated XGBoost, enabling tree‑based models alongside deep‑learning models.

Added support for unsupervised‑learning workflows, allowing clustering analysis within SQLFlow.

Implemented SHAP‑based explanation visualizations for both deep‑learning and tree models.

Provided compatibility with Didi’s Hive data warehouse.

Architecture Vision

The “SQL garden” concept treats SQLFlow as a foundational platform where domain‑specific plug‑ins (e.g., transportation) can be added, creating an open‑source marketplace of reusable SQL‑driven AI solutions.

Resources

Website: https://sqlflow.org

GitHub repository: https://github.com/sql-machine-learning/sqlflow

Docker command to run the car‑price prediction example (XGBoost model with SHAP explanations):

docker run -p 8888:8888 sqlflow/sqlflow:didi