Kalshi
A low-latency Kalshi research pipeline built around a retrained Moonshine speech model.
Co-developed an ultra-low-latency speech recognition and automated trading system for prediction markets, generating ~$10,000 in its first month of deployment. Fine-tuned open-weight speech recognition models and engineered substantial portions of a custom inference stack, optimizing model execution and end-to-end latency down to ~5 ms from audio packet receipt to trade submission, outperforming competing bots by up to 20 ms.
The access code to the technical writeup is unfortunately available only if you ask directly by email due to the extremely competitive nature of these bots.
