TradeEye

Inspiration

We were drawn to prediction markets where everyone sees the same live moment, but price only moves after the crowd reacts. We asked: what if the edge wasn’t a smarter fair-value model, but speed, acting on speech before the rest of the market catches up? Kalshi’s “what will they say?” style markets felt like the right sandbox.

What it does

TradeEye listens to live audio, transcribes it in real time, and uses Gemini to score how likely each tracked word is to be said. The frontend shows a trading-style dashboard: word cards, probabilities, a synthetic order book, and simulated trades so we can see whether faster reads would have paid off. We also surface real past and upcoming Kalshi mention markets.

How we built it

We split backend and frontend. Backend (backend.py): FastAPI, Google Cloud Speech-to-Text for streaming transcription, Gemini for word-level scoring, and WebSockets to push transcript tokens and JSON predictions to connected clients. Frontend: Next.js, Kalshi API integration for mention markets, and a small in-browser trading engine tied to those predictions.

Challenges we ran into

Applying threading to our program, so it can transcribe and send + receive data from Gemini at the same time.

Accomplishments that we're proud of

Proud of working as a team efficiently and learning from everyone's ideas.

What we learned

We learned more about how AI can be integrated and build apps that were not possible without these API models.

What's next for TradeEye

We’d plug into authenticated Kalshi trading for paper or small-size real execution, tighten the model with evals on recorded clips, and add clearer metrics on signal-to-trade latency. A proper replay mode would let us backtest “if we’d fired here, what fills would we have gotten?” without needing a live every time.