-
Our agent just predicted probabilities for NBA playoffs, Premier League, NHL MVP, EuroLeague and more, all in under 20 seconds
-
Our prediction server is up and running on Railway and it'll keep serving forecasts for the next 2 weeks without us touching anything.
What it does
Our agent receives prediction market questions across any domain — sports, economics, politics, entertainment — and returns calibrated probability estimates for each outcome. It combines three data sources: Kalshi real-money market prices as a baseline anchor, Brave web search for fresh live data, and GPT-4o-mini for intelligent reasoning across all domains.
How we built it
Python + FastAPI — HTTP server that accepts prediction requests Kalshi API — fetches real trader prices as probability anchor Brave Search API — searches web for current standings, news, odds OpenRouter + GPT-4o-mini — reasons across any domain Railway — hosts our server publicly for 2-week evaluation Prophet Arena CLI — for local testing and registration
Challenges
Getting the server public and registered within hackathon time Handling JSON parsing failures for events with 30+ outcomes Preventing overconfident predictions (0% or 100%) Making the agent work for unknown future categories
Accomplishments
Brier score of 0.1726 on sample resolved dataset Successfully deployed public endpoint on Railway Agent handles any event category gracefully
Tech Stack
Python, FastAPI, OpenRouter, Kalshi API, Brave Search API, Railway, Prophet Arena
Log in or sign up for Devpost to join the conversation.