Kairos

AI-Driven Prediction Market Intelligence Platform

Kairos is a desktop application that surfaces potential mispricings in prediction markets (primarily Polymarket) by analyzing news signals and correlating them with active markets. It uses advanced NLP, semantic embeddings, and Kelly criterion-based position sizing to generate actionable betting recommendations.

Disclaimer: Kairos provides informational recommendations only. It does not place trades. Always verify sources and do your own research before allocating capital.

How It Works

1. Data Aggregation: Fetches articles from 40+ sources including Hacker News, NewsAPI, GDELT, and major news RSS feeds (Reuters, Bloomberg, BBC, NYT, etc.)
2. Market Discovery: Retrieves active Polymarket markets with quality filtering (liquidity, volume thresholds)
3. Semantic Matching: Uses OpenAI embeddings to match articles with relevant markets via cosine similarity
4. AI Analysis: GPT-4o-mini analyzes article-market pairs to estimate probabilities and identify edge
5. Risk Assessment: Classifies recommendations into low/medium/high risk tiers based on liquidity, time horizon, and edge magnitude
6. Position Sizing: Applies Kelly criterion for optimal bet sizing within each risk tier

Tech Stack

Backend (Rust + Tauri v2)

• Tauri v2: Cross-platform desktop framework
• Tokio: Async runtime for concurrent API fetching
• Reqwest: HTTP client for news and market APIs
• Async-OpenAI: Embeddings and LLM analysis
• Feed-rs: RSS/Atom feed parsing
• Rusqlite: Local caching of articles and markets
• Regex: LLM response parsing

Frontend (Svelte + TypeScript)

• Svelte 5: Reactive UI framework
• TailwindCSS: Utility-first styling
• Vite: Build tool and dev server

Getting Started

Prerequisites

• Rust (latest stable)
• Node.js (LTS)
• pnpm (recommended)
• OpenAI API key (required for embeddings + LLM analysis)
• NewsAPI key (optional, for additional news coverage)

Installation

1. Clone the repository:

   git clone https://github.com/yourusername/kairos.git
   cd kairos

2. Install frontend dependencies:

   pnpm install

3. Configure environment variables:

   # src-tauri/.env
   OPENAI_API_KEY=your_key_here
   NEWSAPI_KEY=your_key_here

4. Run in development mode:

   pnpm tauri dev

Features

Implemented

• Multi-source news aggregation (40+ RSS feeds, Hacker News, NewsAPI, GDELT)
• Polymarket integration with quality filtering
• Semantic embeddings for article-market matching (text-embedding-3-small)
• LLM-powered analysis with structured probability estimates
• Risk-tiered recommendations (low/medium/high)
• Kelly criterion position sizing with fractional Kelly for risk management
• Topic clustering to reduce correlated exposure
• Source weighting (Reuters, Bloomberg, AP get higher weights)
• Recency decay for article relevance
• Progress streaming to UI during analysis
• Local SQLite caching for articles and markets

Analysis Pipeline

1. Fetch Phase: Retrieves up to 200 HN stories, 500 NewsAPI articles, and 40+ RSS feeds
2. Filter Phase: Deduplicates and filters articles; filters markets by liquidity (> $5k) and volume (> $10k)
3. Embedding Phase: Generates embeddings for top 150 articles and up to 1000 markets
4. Matching Phase: Cosine similarity + keyword overlap scoring
5. LLM Analysis Phase: Deep analysis of top 25 article-market matches
6. Clustering Phase: Groups similar markets to limit correlated bets
7. Allocation Phase: Kelly-based budget allocation across risk tiers

Configuration

Environment Variables

Variable	Required	Description
OPENAI_API_KEY	Yes	For embeddings and LLM analysis
NEWSAPI_KEY	No	For additional news coverage
ECON_CALENDAR_URL	No	External economic calendar RSS
POLL_FEED_URL	No	External polling data feed

Risk Tiers

Tier	Budget Allocation	Characteristics
Low Risk	50-70%	High liquidity, near-term resolution, modest edge
Medium Risk	20-30%	Moderate liquidity, medium-term, solid edge
High Risk	10-20%	Lower liquidity, long-term, high edge

Architecture

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│   News Sources  │────▶│  Rust Backend   │────▶│  OpenAI APIs    │
│  (HN/API/RSS)   │     │  (Tauri/Core)   │     │ (Embeddings+LLM)│
└─────────────────┘     └─────────────────┘     └─────────────────┘
                               │
                               ▼
                        ┌─────────────────┐
                        │  SQLite Cache   │
                        └─────────────────┘
                               │
                               ▼
┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│  Svelte Frontend│◀────│  Analysis Engine│◀────│ Polymarket API  │
│   (UI/Results)  │     │ (Matching/Risk) │     │  (Markets)      │
└─────────────────┘     └─────────────────┘     └─────────────────┘

Current Limitations

• Article content is limited to titles and summaries (full text extraction not implemented)
• No historical backtesting or performance tracking yet
• Polymarket API doesn't provide historical price data for momentum analysis
• LLM analysis limited to 25 top matches to manage API costs
• No automated trade execution (manual only)

Future Roadmap

• Full article text extraction for deeper analysis
• Historical performance tracking and calibration
• Additional prediction markets (Kalshi, etc.)
• Momentum and technical indicators
• Custom user-defined watchlists
• Export recommendations to CSV
• Confidence calibration based on historical accuracy

License

Distributed under the MIT License.