The team developed an AI-powered explanation feature that effectively translates complex, multi-program policy calculations into clear and accessible explanations, enabling users to explore "what-if" scenarios and understand key factors influencing benefit amounts and eligibility thresholds.
The team conducted experiments to determine whether clients would be responsive to proactive support offered by a chatbot, and identify the ideal timing of the intervention.
At Rules as Code Demo Day we heard from Song Hia of the NYC Mayor’s Office for Economic Opportunity and Ethan Lo of the NYC Office of Technology and Innovation who demoed the NYC Benefits Platform Screening API which provides machine-readable calculations and criteria for benefits screening that power the ACCESS NYC screening questionnaire. This makes it easier for NYC residents to discover multiple benefits they may be eligible for. The City is now extending the API to support the new MyCity platform, a one-stop shop for all services and benefits.
We wrapped up Rules as Code Demo Day with Max Ghenis and Nikhil Woodruff, the founders of PolicyEngine. The PolicyEngine web app computes the impact of tax and benefit policy in the US and the UK. With PolicyEngine, anyone can freely calculate their taxes and benefits under current law and customizable policy reforms, and also estimate the society-wide impacts of those reforms. Policymakers and think tanks from across the political spectrum can analyze actual policy. PolicyEngine is built atop the open source OpenFisca US and UK microsimulation models and they are building an open unified data set utilizing data from the Policy Rules Database, Current Population Survey, Survey of Consumer Finances, Consumer Expenditures, tax records, and IRS Public Use File.
We continued Rules as Code Demo Day with Daniel Singer and Preston Cabe from Benefits Data Trust. Benefits Data Trust provides benefit outreach and application assistance services in seven states. Using Benefits Launch, their in-house interview and rules engine, they support two hundred contact center employees as they screen and apply thousands of clients each year. They also offer a self-service screener, Benefits Launch Express. Additionally, they offer an eligibility API to integrate with other services.
The team explored using LLMs to interpret the Program Operations Manual System (POMS) into plain language logic models and flowcharts as educational resources for SSI and SSDI eligibility, benchmarking LLMs in RAG methods for reliability in answering queries and providing useful instructions to users.
Building on our February 2022 report Benefit Eligibility Rules as Code: Reducing the Gap Between Policy and Service Delivery for the Safety Net, the Beeck Center’s Digital Benefits Network (DBN) hosted Rules as Code Demo Day on June 28, 2022 where there were eight demonstrations of projects and code followed by a collaborative problem solving session on how to continue advancing rules as code for the U.S. social safety net.
The team explored the performance of various AI chatbots and LLMs in supporting the adoption of Rules as Code for SNAP and Medicaid policies using policy data from Georgia and Oklahoma.
At Rules as Code Demo Day Executive Director Zareena Mayn and Chief Technology Officer Dize Hacioglu of mRelief demoed the code for their Supplemental Nutrition Assistance Program (SNAP) eligibility screener. mRelief is a women-led team that provides a web-based and text message-based SNAP eligibility screener to all 53 states and territories that participate in SNAP. They demonstrated how they have modularized their code to host federal program rules and state-specific rules.
The team developed an application to simplify Medicaid and CHIP applications through LLM APIs while addressing limitations such as hallucinations and outdated information by implementing a selective input process for clean and current data.
The team examined how AI, specifically LLMs, could streamline the case review process for SNAP applications to alleviate the burden on case workers while potentially improving accuracy.
