AI-Augmented Deliberative Committee

Multi-agent deliberation system using Claude Opus/Haiku with 7 expert jurors and 8 demographically-grounded community stakeholders (ACS 2022) to evaluate Chicago urban policy proposals

Overview

The AI-Augmented Deliberative Committee addresses a fundamental governance challenge: how do you incorporate diverse voices when convening 2.7 million Chicagoans in one room is not feasible? The system simulates a deliberative panel of expert jurors and community stakeholders to evaluate stadium deals, public subsidies, and urban development proposals through structured multi-round deliberation.

What It Does

The system orchestrates two types of AI agents through a three-round deliberation protocol:

Expert Jury (7 members):

  • Public Finance & Municipal Governance Expert
  • Urban Economist
  • Community Organizer
  • Sports & Entertainment Industry Analyst
  • Environmental Sustainability Expert
  • Political Strategist & Legislative Affairs
  • Youth & Education Policy Specialist

Community Stakeholders (8 “digital doubles”): Demographically representative personas grounded in ACS 2022 5-year census data across Chicago neighborhoods (Pilsen, Bronzeville, Austin, Rogers Park, South Loop, Lincoln Park, Englewood, Near West Side). Each persona is built from real census tables: income (B19013), housing tenure (B25003), commute patterns (B08301), education (B15003), and occupation (C24010).

Three-Round Protocol:

  1. Round 1: Individual scoring on Impact, Fiscal Responsibility, and Sustainability (1-10 scale) with written justifications
  2. Round 2: Agents see each other’s scores and deliberate in character, surfacing disagreements and tradeoffs
  3. Round 3: Final verdicts and a synthesis report capturing consensus, tensions, and conditions for recommendation

Technical Architecture

Two-Tier LLM Design:

  • Claude Opus: Summarizes proposals up to 200K characters into structured summaries
  • Claude Haiku: Cost-efficient deliberation across all three rounds

Key Engineering Decisions:

  • Anthropic structured outputs (JSON schema) enforce valid scoring format in Rounds 1 and 3
  • Conversation history passed between rounds to maintain deliberation context
  • Regex fallback parsing if JSON extraction fails
  • Rate limiting (2.5s delays) to stay under API limits
  • Context windowing: 200K chars for summarization, 120K for deliberation

Technologies & Skills

  • Python with Anthropic SDK for LLM orchestration
  • Streamlit for interactive web interface
  • Docker & Docker Compose for containerization
  • UV package manager for dependency management
  • pytest with integration test markers
  • Ruff for linting and formatting
  • Jupyter notebooks for demo and analysis

Team & My Role

HPIC AI Challenge submission (Harris Public Policy Challenge, March 2026)

My contributions: All code, architecture, agent design, deployment, evaluation suite, and documentation. The original team created the policy proposal that the system evaluates.

Source Code

Repository: anfelipecb/AI-Augmetnted-Deliberative-Committee

Live Application: Streamlit App