TED Talk Companion

The NAID research agent was developed in Visual Studio Code using Anthropic's Claude API, configured as a custom AI agent with a structured system prompt defining its analytical scope, data access protocols, and response standards. The agent is grounded in four proprietary datasets — a GTAP-based county-level U.S. labor simulation database (covering baseline employment and five policy scenarios), a Mexican-origin and Latino diaspora GDP contribution dataset at the state level, a Mexico export-linked employment dataset by sector and U.S. destination state, and a remittances dataset tracking quarterly flows by U.S. state of origin — all serialized as Parquet files and made accessible to the agent's sandboxed code execution environment at runtime. When a user submits a query, the agent dynamically loads and queries the relevant dataset(s) using Python (pandas), performs aggregations and filtering in-session, and returns policy-grade outputs including tables, prose analysis, and matplotlib visualizations. For questions requiring current information beyond the static datasets — such as recent policy announcements or new economic data — the agent invokes a web search tool and applies strict citation discipline. All numeric claims are traced back to the source dataset, scenario, and vintage, with caveats surfaced proactively where model limitations apply (e.g., GTAP wage artifacts under extreme deportation scenarios). The system prompt enforces an audience-appropriate style — precise, non-advocacy, and structured for use in researcher briefings and government policy offices.

Created by Adhila Akbar, NAID Researcher