1 of 13

CivicMind: AI-Powered Civic Debate & Policy Simulation

PRESENTED BY: MAHARSHI JANI

2 of 13

Problem Statement

  • Public policy often lacks real-time citizen feedback and fails to predict social impact.
  • Traditional simulations are slow, expensive, and not scalable.
  • There’s no interactive, AI-based platform to simulate debates on policies with diverse civic viewpoints.
  • Policymakers lack access to simulated civic feedback loops before implementation, increasing the risk of policy backlash.
  • Existing tools are either data-heavy dashboards or static surveys, not interactive or conversational.

3 of 13

Project Objective

  • To build CivicMind, an interactive platform where multi-agent AI simulates public discourse, evaluates policies, and generates structured civic insights.

Key goals:

  • Simulate diverse civic personas debating public policies.
  • Use LLM agents to reason, debate, and suggest improvements.
  • Provide policymakers and citizens with AI-powered feedback reports.

​

4 of 13

Key Features

  • 🤖 AI Persona Agents: Simulated stakeholders (Politician, Citizen, Activist, Business, Journalist, Custom) provide diverse viewpoints.
  • 🧠 Multi-Agent Debate Engine: AI personas with memory, tone, and reasoning.
  • 📄 Policy Upload & Summarization: Accepts real-world policies (PDF/text).
  • 💬 Dynamic Tone System: Agents evolve their stance after each round.
  • 📈 Sentiment Analysis: Tracks how agents emotionally perceive the policy.
  • ⚖️ JudgeBot: Evaluates feasibility, sentiment, impact, and offers recommendations.
  • 📊 Interactive Streamlit UI for real-time visualization.
  • 📂 Session Memory & Logs: All debates, agent responses, and evaluations are stored for review, comparison, and longitudinal policy analysis.

5 of 13

System Architecture

  1. User Uploads Policy
  2. LLM Summarizer → generates a concise summary
  3. Persona Generator → creates diverse civic agents
  4. Debate Simulator → agents exchange arguments over rounds
  5. Tone Recalibrator → updates agent tones dynamically
  6. JudgeBot → evaluates policy & generates report

6 of 13

Technologies Used

  • LLMs: Groq (LLM-based reasoning and debate)
  • Streamlit: Frontend for user interaction
  • PDF/Text Parsing: PyPDF2
  • Matplotlib/Pandas: Sentiment Visualization
  • Python 3.11

7 of 13

Multi-Agent Simulation

  • Each agent has:
    • Persona (e.g., student, senior citizen, small business owner)
    • Belief system (liberal, conservative, neutral)
    • Memory and tone (supportive, skeptical, aggressive)
  • Agents debate over 2–3 rounds
  • Tone adjusts based on prior arguments (via custom logic)

8 of 13

JudgeBot Evaluation Report

  • Uses LLM + rules-based prompts for structured output
  • Components:
    • 🧪 Feasibility Score
    • 🗳️ Public Sentiment Analysis
    • 🌍 Predicted Social Impact
    • 🧭 Policy Recommendations

9 of 13

Real-World Use Cases

  • Civic education platforms
  • Government policy testing sandbox
  • NGOs gathering public feedback
  • Journalism & media analysis
  • Classroom simulations for political science

10 of 13

Challenges Faced

  • Ensuring consistent agent personas across turns
  • Balancing randomness vs realism in agent responses
  • Handling hallucination and bias in LLM outputs
  • Prompt tuning for JudgeBot’s structured feedback

11 of 13

Future Work

  • User-as-Agent Participation: Let users join debates in real-time to simulate hybrid civic engagement
  • Integrate RAG with real-time news/civic databases
  • Multilingual agent support
  • RL-based policy improvement agents

12 of 13

Conclusion

“Bridging Governance and AI – One Debate at a Time.”

  • CivicMind brings AI-driven civic intelligence to the forefront.
  • Can empower governments, citizens, and researchers alike.
  • Enhances transparency and trust in public policy decisions.
  • Promotes inclusive governance by simulating diverse perspectives.
  • Offers a scalable tool for civic education and engagement.

13 of 13

Thank You