Inspiration
Societies are shaped by public policies, which are frequently implemented without thorough simulations or open audits. The increasing demand for civic technologies that are proactive as well as reactive served as our inspiration. We saw a chance to create a system that gives decision-makers insight, foresight, and fairness with the emergence of AI and multi-agent systems.
The concept for PolisMind was inspired by the idea of a civic advisor who could independently read policies, model their effects on society, optimize them, and then justify their decisions. Current policy simulators lack inter-agent intelligence, are manual, or have a limited scope. Our goal was to alter that.
What it does
Using the Agent Development Kit (ADK), PolisMind is a multi-agent policy auditor and simulator. It uses Big Query to parse real-world policy documents and map them to national or regional datasets.
It models how the policy will affect different stakeholders in the short, medium, and long terms. Examines possible environmental and socioeconomic effects
- Uses agent collaboration to optimize the policy The results are presented in a report that is readable by humans and includes a dynamic visualization.
For legislators, non-governmental organizations, and civic technologists, this establishes an open, data-supported, and intelligent feedback loop.
How we built it
Using the Python version of ADK, we constructed PolisMind with an emphasis on the specialization and orchestration of the following agents:
- Policy Parser Agent – This tool extracts structured intent from policy documents using Vertex AI and LLMs.
- Data Mapping Agent - Looks for pertinent datasets related to the policy domain by querying Google Big Query.
- Simulation Agent - Predicts ripple effects across demographics by running simulations using agent-based modeling. Using unique scoring logic, the Impact Evaluator Agent assigns scores to ethical, economic, and environmental impacts.
- Optimization Agent - Makes recommendations for enhancements through heuristic optimization and reinforcement learning.
- Narrator Agent - Creates detailed reports for different stakeholders using the Gemini API.
We utilized Agent Engine for lifecycle management, stored and queried datasets using Big Query, and deployed the agent backend on Cloud Run.
Challenges we ran into
Complexity of data integration: Converting unstructured policies to structured data was difficult and necessitated the development of tools for semantic alignment. Inter-agent communication: It was difficult and took a number of iterations to design protocols that would allow agents to coordinate without redundancy. Simulation modeling: It was difficult to create quick and realistic simulations that captured the subtleties of how policies affected different socioeconomic groups. LLM hallucination: Filtering and reinforcement were necessary to ensure factual reliability when summarizing or producing insights.
Accomplishments that we're proud of
-Using ADK, a completely modular, multi-agent system was designed and put into operation. -Using real-world data sources that were directly connected to the parsed policy intent, dynamic simulations were constructed.
- Six specialized agents working together were orchestrated seamlessly.
- Produced policy audit reports that are actionable and readable by non-technical stakeholders.
What we learned
-How to use ADK to orchestrate cooperative agent systems with asynchronous logic and clearly defined boundaries. -How to incorporate Big Query and Vertex AI, two Google Cloud tools, and AI models into multi-agent workflows.
- The complexity of computational policy modeling and the subtleties of civic policy language. When agents collaborate, they can reveal insights that LLMs by themselves are unable to.
What's next for PolisMind
- Open civic platform: Make PolisMind a civic intelligence tool available to citizens, NGOs, and governments. -Simulation of geography: Connect to the Google Maps API to map the effects of policies by location. -Explainable AI layer: Create an open record of each choice PolisMind makes or suggestion it makes.
- Policy versioning: Include the ability to simulate delta effects and compare policy drafts. -Public beta: Introduce a streamlined, open-access version that allows users to upload a policy and receive an audit immediately.
We think PolisMind has the potential to be a key component of AI-guided governance — an accountable, transparent, and data-driven approach.
Log in or sign up for Devpost to join the conversation.