TrueOdds.AI

• 

  TrueOdds.AI Logo

TrueOdds.AI was build during ProphetHacks 2026 by Antonio Unabia, Daniel Danque, Emilio Calvo, and Steve Nuevaorlanda.

Inspiration Sports and prediction markets have always fascinated us, specifically the idea that you can quantify uncertainty and assign a number to the likelihood of any outcome. When we discovered the Prophet Arena benchmark, we saw the perfect opportunity to combine that interest with AI. We wanted to build something that could reason about real-world events the way a sharp analyst would, by weighing evidence, considering context, and arriving at a calibrated probability rather than a gut feeling.

What it does TrueOdds.AI is an AI-powered forecasting agent that takes real-world binary-outcome events across sports, economics, and entertainment and returns calibrated probability estimates for every possible outcome. Rather than simply predicting yes or no, it reasons across the full field of outcomes simultaneously, accounting for current standings, recent form, historical base rates, and live information, then normalizes the probabilities into a coherent prediction.

How we built it We built TrueOdds.AI using the ai-prophet-core SDK and CLI as the backbone for event retrieval, prediction formatting, and scoring. Our agent is written in Python and calls Perplexity Sonar Pro via OpenRouter, which gives us access to real-time web search baked directly into the model's reasoning process. The agent is structured as both a local Python module for testing and a FastAPI HTTP endpoint for live judge evaluation. We also integrated the Odds API to supplement the model's context with current market data, giving it an edge on sports events where real-time information matters most.

Challenges we ran into This was our entire team's first time developing an AI-powered application, so the learning curve was steep from day one. Getting the environment set up correctly across different machines bt managing virtual environments, Python path issues, Windows encoding errors, and API keys, took significant time before we even wrote a single line of agent logic. We also ran into challenges around the output format, since the CLI expected a specific response schema and silent mismatches caused predictions to be skipped without obvious error messages. Understanding the difference between local module mode and HTTP endpoint mode, and making our agent compatible with both simultaneously, was another hurdle we had to work through carefully.

Accomplishments that we're proud of For a team with no prior experience building AI models, getting a fully functional forecasting agent running end-to-end, from event retrieval to live predictions to a deployed HTTP endpoint, in a hackathon timeframe is something we're genuinely proud of. We're also proud of how modular and clean the final architecture is. Swapping models, adjusting prompting strategies, or pointing the agent at a different dataset requires minimal changes, which speaks to how thoughtfully we structured the codebase as we went.

What we learned We learned an enormous amount about how LLMs work in practice, not just how to call an API, but how to engineer prompts that produce consistent, structured outputs and how to handle edge cases when the model doesn't respond exactly as expected. We also learned how important calibration is in forecasting: a model that always says 90% confident is far less useful than one that accurately reflects its uncertainty. On the technical side, we got hands-on experience with FastAPI, virtual environments, git workflows, and integrating multiple external APIs into a single pipeline.

What's next for TrueOdds.AI We want to expand TrueOdds.AI beyond sports into economics and political events, where calibrated forecasting has real-world impact. We're also exploring fine-tuning a smaller open-source model on historical prediction market data to reduce reliance on large commercial APIs. Longer term, we'd love to build a leaderboard-facing dashboard that visualizes our predictions and accuracy over time, making TrueOdds.AI not just a benchmark agent but a fully transparent forecasting platform.

Built With

• fastapi
• openrouter
• python
• render