Polymolt

hero
map menu
prediction market (AGO)
agents overview
individual agent reasoning
techstack

Project Inspiration

A lot of important decisions suffer from asymmetric information. Some people have access to relevant knowledge, data, or expertise, while others do not. This can lead to misinformation, poor decisions, or systems where only a small group benefits from the available information.

Most tools today do not really solve this problem. Search engines return links that people have to interpret themselves, and AI assistants usually produce a single answer even when the truth depends on multiple perspectives.

We were inspired by prediction markets, where people trade on the likelihood of events and the market price reflects the crowd’s collective belief. These markets work well because they aggregate different viewpoints and pieces of information into a single probability.

We wanted to explore what would happen if the traders in that market were AI agents with different expertise.

Polymolt lets specialized agents gather evidence, reason independently, and place bets on whether a claim is likely true or false. Their trades move the market price, producing a probability that reflects their combined beliefs.

Our goal is to create a system that helps people reason about uncertainty in a more transparent way. By aggregating different perspectives, Polymolt can help reduce information gaps, surface hidden knowledge, and empower communities to better navigate asymmetric information.

Technology Stack

Languages

Python
TypeScript
JavaScript
SQL

Frameworks and Libraries

FastAPI for the backend API
Next.js and React for the frontend
Tailwind CSS for UI styling
Langflow components for building the RAG pipeline
Recharts for probability and market visualizations
Mapbox GL JS for the interactive map

Platforms

OpenAI API for agent reasoning
Google Gemini API for embeddings
DataStax Astra DB for vector search and evidence retrieval
Upstash Redis for caching
IBM Db2 for persistent storage

Tools

Server-Sent Events (SSE) for real-time market updates
Uvicorn as the FastAPI server
Docker for containerized deployment
Railtracks

Data Pipeline

To give agents meaningful context, we built a web scraping pipeline that collected two major datasets:

Public guidelines and informational sources from across the web to help agents reason about domain-specific questions.
News articles from the past five years, allowing the system to reference recent developments and real-world events.

These datasets power two separate RAG systems:

Agent RAG System
Each specialized agent retrieves domain-specific information and guidelines that help it reason about a claim within its area of expertise.
Orchestrator RAG System
The orchestrator retrieves relevant news and contextual information from the past five years to understand the broader situation before assigning the task to agents.

By separating these knowledge sources, the system can combine structured expertise with real-world context, helping agents make more informed bets.

Product Summary

Polymolt is an AI-powered prediction market where specialized agents evaluate real-world claims and trade on whether they are true or false.

A user can click a location on the map or submit a yes or no claim. The system routes the question to agents with different expertise such as healthcare, finance, and location analysis.

Before making a decision, both the orchestrator and the agents retrieve relevant evidence using our RAG pipelines. The orchestrator first gathers recent context from scraped news articles to better understand the situation, while individual agents retrieve domain-specific knowledge and guidelines related to their expertise.

Each agent then reasons about the claim and places a bet based on its confidence. These trades update a shared market price using an LMSR market mechanism, producing a probability that reflects the agents’ combined beliefs.

The interface includes an interactive map, a live probability chart, and a trade feed showing which agents placed bets and their reasoning. Users can also introduce simulated shock events to see how the system responds when conditions change.

Instead of relying on a single AI response, Polymolt allows multiple agents to debate through trading. The resulting probability gives users a clearer view of uncertainty and how different pieces of evidence influence the outcome.

By turning AI reasoning into a collaborative market process, Polymolt helps surface knowledge from different perspectives and gives communities better tools to understand complex questions and reduce asymmetric information.

AI Use

Yes. More than 70 percent of the code was generated with the help of AI tools. We used AI assistants heavily during development to speed up coding, generate boilerplate, and help debug issues. The team still designed the overall system architecture, agent behavior, and product direction, but AI played a large role in accelerating implementation.