DataDirector
DataDirector is an applied-AI voter data platform for field organizers, regional directors, and campaign managers. I built the system so campaigns can ask multi-step questions in plain language instead of waiting on a data team to assemble SQL queries and spreadsheets.
Technologies used
The problem
Voter files are wide, large, and difficult to explore without technical support. A useful answer may require filtering multiple fields, comparing groups, counting records, or building a cross-tabulation across millions of rows. Sending the full dataset to a language model is neither practical nor reliable.
What I built
- A multi-tool AI agent with 15+ specialized read-only tools
- Natural-language analysis across 4.2M+ voter records, 197 fields, and 3 states
- Multi-step query planning for filters, counts, comparisons, and cross-tabulations
- Intent classification that routes questions to the right deterministic data operation
- Streaming responses and charts for results that are easier to inspect
- A production stack using Next.js, FastAPI, PostgreSQL, Supabase, and Vercel
Engineering approach
The language model interprets the request and selects constrained tools; the tools perform the actual data operations. This keeps database access controlled and makes the result easier to validate than an answer generated directly from raw records.
I also built the public product experience and the interfaces used to present query results, charts, and campaign-oriented assistant personas.
Stack
- Next.js and React
- Python and FastAPI
- PostgreSQL and Supabase
- Vercel
- Constrained AI tool calling
Status
Ongoing product work since 2025.