After watching Glitch wind down and spending three months discovering Iām definitely not a āgrowth teamā person at Handshake, I found myself back where I belong: civic tech. Three months trying to make venture-backed growth metrics move taught me that what I actually want is to build things that help people navigate government services. So when I heard about an AI Residency at Propel focused on the social safety net, I signed up.
Propel is an app that helps people check their EBT balance, find stores that accept benefits, and get updates about their stateās programs. About 5 million Americans use it, roughly 1 in 4 SNAP beneficiaries. The team there has built something genuinely helpful, and theyāve earned deep trust with their users by actually giving a damn about making government benefits easier to navigate.
But I didnāt join to work on the main app. I joined a new residency program specifically focused on HR1 and its impact on SNAP and Medicaid. Weāre exploring how AI might help states and beneficiaries deal with the massive changes coming down the pipeline.
What HR1 Actually Does (abridged)
Hereās what HR1 does: it adds new purchase restrictions and work requirements for SNAP beneficiaries while increasing administrative costs for states. And this is the part that has states panicking: it penalizes states financially based on their error rates. The higher your error rate in administering benefits, the more you have to pay to run the program. States are scrambling to understand their current error rates and figure out how to reduce them before these penalties kick in.
The Structure of the Residency
Our residency pairs three engineers with three Subject Matter Experts whoāve actually run these programs. These folks have been state administrators, theyāve implemented SNAP and Medicaid, theyāve sat in the rooms where these decisions get made. Our role as engineers is to support their vision and use their connections to build what they know states need.
The goal is to demonstrate what responsible AI actually looks like in social safety net programs, and give states a model that isnāt just a vendor making bold promises about āsolving poverty with machine learningā.
Real Work During the Shutdown
During the recent government shutdown, I got to see what this looks like in practice. I built crawlers to pull information directly from state websites and get it in front of people through Propel. Usage tripled at peak. People were checking constantly for any updates about their benefits. The food stamps subreddit had screenshots from Propel all over it, with people sharing what theyād learned with each other.
I used Jina.ai to build what I believe is the most comprehensive database of active food pantries and banks. This tool turns messy web pages into clean markdown with preserved links. Work that would have taken weeks of manual data entry five years ago now takes hours. Iām still skeptical about LLMs broadly, but for specific tasks like āparse these 500 county websites and extract food pantry hours,ā they work.
What weāre building
States are being forced to move quickly, which creates both opportunity and risk. Thereās real appetite for new approaches, but when government moves fast, vendors appear with promises of AI magic thatāll solve everything.
The practical work: systems that help caseworkers spot error patterns before they become expensive penalties; crawlers monitoring how states implement policy changes so they can learn from each other. On the beneficiary side, Iām building tools that translate federal requirements into clear implementation guides and verification flows that actually work for people with complicated lives.
This list could look different in two weeks, and thatās fine.
Why Propelās position is unusual here
States have to act. HR1 isnāt optional, and the penalty structure means slow movers will pay for it. What Propel has that most vendors donāt is an existing feedback loop: we can see how policy changes land for 5 million people, often before states fully understand whatās happening. Thatās a different starting point than typical government modernization work, where the builder and the affected user are completely disconnected.
If youāre working on AI and HR1 responses from any direction (state agencies, cities trying to fill gaps, contractors who care about outcomes), please reach out.