In a move that merges fiscal transparency with electoral participation, SerenDB has deployed an AI agent named Caroline that constructs customized voter education guides by analyzing users' personal financial records. Published on Hacker News on November 3, 2026, the tool requires users to upload tax returns or connect cloud storage services like Google Drive, Microsoft, or OneDrive. Unlike generic voting apps, Caroline parses several years of income data and home ownership documents, such as purchase contracts and property tax bills, to determine what is at stake for the specific household.

Methodology and Data Sources

Caroline operates by identifying every contest on a user's specific ballot through official sources and providing one clear recommendation per race. The agent bases its picks on a rigorous analysis of official fiscal analyses for state and local measures, candidates' voting records, and judges' public discipline records. It also evaluates linked measures as a cohesive plan and checks claims from campaign mailers against the historical record. For local races, including those using ranked-choice voting, the agent provides full rankings, while also tracking money moves tied to the ballot, such as deadlines and actions with attached dollar amounts.

Privacy and Cost Structure

The system claims strict nonpartisanship, asserting that a single method applies to every candidate and party, with decisions driven solely by the user's records and priorities rather than the AI's internal views. Privacy is handled by keeping documents within the user's Seren account, ensuring they are used only for guide generation. The agent's web searches are anonymized, never including the user's name, address, or income. Users can request the erasure of their records at any time, and firing the agent deletes all uploads and chat history. The service charges only the actual compute cost of running the guide, with no additional fees, averaging $0.63 per completed task.

Key Takeaways

  • Caroline uses tax returns and home papers to calculate the direct dollar impact of ballot measures on a household.
  • The agent provides one clear pick per contest, backed by dated sources and fiscal analyses.
  • Privacy is prioritized with anonymized web searches and user-controlled data erasure options.
  • The service costs approximately $0.63 per guide, with no hidden subscription fees.

The Bottom Line

This is a fascinating experiment in personalized civic tech, but relying on tax data to predict political preferences is a bold, potentially reductive move that assumes economic self-interest is the sole driver of voter behavior. The integration of financial records into voting advice creates a hyper-personalized experience that standard apps cannot match, though the low HN engagement suggests the market is still skeptical of AI agents handling sensitive personal finance data.