ChatDCA
An AI assistant for New Jersey state staff.
Overview
State staff ask one assistant. It hands the work to the right specialist.
ChatDCA is a team of five agents. A front-door assistant takes the question and passes it to a specialist for data analysis, forms or writing. Staff see one conversation either way.
This page is short on purpose. It's government work, so most of the details stay with the team. What I can share is how it's put together.
Three decisions
Build on a framework, and fix it upstream when it breaks.
ChatDCA runs on eve, Vercel's agent framework. When something in eve didn't hold up for us, the fix went back to eve as a pull request instead of living in a private patch.
Every model call goes through AWS Bedrock.
The agents reach their models through Bedrock, so the whole system stays with one cloud provider the department already uses.
Code an agent writes runs in a sandbox, never next to the app.
When the data agent writes code to answer a question, that code runs in an isolated sandbox, so a bad script can only break itself.