Sports aren’t just about the score. They’re about momentum, emotion, and narrative. A 10-point lead in the first quarter means nothing, but the same lead with two minutes left feels inevitable. Most AI systems treat games as static data streams. We wanted to build something that actually feels the game.
That’s where this project started: could an AI watch a live game, understand what’s happening in context, and respond the way a real fan would, instantly, emotionally, and with personality?
We built an agentic system that monitors live sports data, interprets game state (score, time, momentum shifts, clutch moments), and generates contextual trash talk in real time. Instead of a single prompt-response model, the system follows a loop: it decides what information it needs, fetches it through APIs, evaluates the situation, and then produces a response calibrated to the moment. A blowout, a comeback, and a last-second collapse all trigger fundamentally different tones.
One of the biggest challenges was defining “momentum” in a way that actually reflects how people perceive games. Raw score differences weren’t enough. We had to factor in time remaining and scoring runs.
A key constraint we had to work around was cost and efficiency. External APIs and language models aren’t free, and naïve implementations can burn through tokens extremely quickly. That forced us to be intentional about every call the system makes. We reduced unnecessary API requests, structured context to be as compact as possible, and used lightweight models when full reasoning wasn’t needed.
Built With
- agentic
- api
- flask
- html
- javascript
- json
- python
Log in or sign up for Devpost to join the conversation.