June 17, 2025
Estimand
Fall 2023
- Causal AI Architecture & Design: Conceptualized and implemented a novel causal inference system tailored for time series analysis, incorporating counterfactual reasoning with dynamic temporal variables.
- Algorithm Development: Invented a new method for applying counterfactual logic to time series data by extending the DoWhy framework with support for Dynamic Bayesian Networks (DBNs) and temporal interventions.
- Framework Extension: Deeply modified and enhanced the internals of the DoWhy library to support time-aware counterfactual queries and causal graphs that evolve over time.
- Full-Stack ML Development: Built the MVP end-to-end—from data ingestion and preprocessing, to modeling, to API/demonstration tooling—with production-level reliability.
- Interactive Demo & Visualization: Created an investor-facing demo exploring causal relationships in macroeconomic data, with capabilities for “freezing” variables and simulating alternate futures.
- Deployment & Operationalization: Successfully deployed causal inference models in a scalable environment, overcoming significant limitations in existing causal libraries not designed for production use.
The project was commissioned with a visionary objective: to create a unique causal AI demonstration that focuses on the causal analysis of time series data. The ultimate goal was to showcase this innovative capability to investors as a proof of concept, highlighting its potential applications and effectiveness in real-world scenarios.As the lead on this project, I undertook the comprehensive task of developing the entire MVP from scratch.
My contributions spanned the entire lifecycle of the project, from conceptualization to implementation. Key among my innovations was the development of a novel algorithm designed to apply counterfactual reasoning to time series data. This required a significant modification and extension of the existing capabilities of the DoWhy package, specifically adapting its counterfactual analysis features to work with dynamic Bayesian networks.
I ventured into uncharted territory by operationalizing causal models in a production environment, a feat not previously achieved with the DoWhy framework. This involved not just algorithmic innovation but also overcoming technical challenges related to deploying these complex models in a scalable and reliable manner.
The demonstration centered around a compelling example: a temporal causal graph that elucidates the relationships between the federal interest rate, mortgage rate, median housing sales price in the United States, and the rental vacancy proportion. This example was carefully chosen to illustrate the intricate interplay of economic factors over time and showcase the model's capability to perform counterfactual analysis by "freezing" variables at specific points in time, allowing for the exploration of alternate scenarios.
The demo also featured the ability to forecast future states based on diverging paths from a set point of divergence, providing a dynamic tool for understanding potential future trends.
The project met and exceeded the client's objectives, culminating in a powerful demonstration of causal AI's potential to untangle complex temporal relationships. The unique approach to integrating counterfactual analysis with time series data garnered significant interest from investors, showcasing the practical applications and the innovative edge of the technology. The successful deployment of causal models in production environments further highlighted the project's technical viability and readiness for real-world application.
The investor interest sparked by this demonstration has poised the client to transition from concept to project development, marking a significant milestone in their journey towards leveraging causal AI for impactful real-world applications. This project stands as a testament to the transformative potential of causal AI in elucidating the dynamics of time series data, paving the way for new insights and advancements across various domains.