Real-time Earnings Calls Pipeline

A real-time pipeline that transcribes earnings calls to text and makes trades based on hard and soft drivers extracted from the audio. An ASR model was fine-tuned on 5,000 hours of earnings call audio to handle financial jargon like EBITDA, guidance, and ticker symbols with high accuracy.

The transcript feeds into a custom sentiment LLM trained to extract both hard drivers (revenue figures, guidance numbers) and soft drivers (management tone, confidence signals). A downstream trading module maps these signals to position-sizing logic and can route orders via a brokerage API.

The project demonstrates how chaining ASR, NLP, and quantitative finance into a single streaming pipeline can surface tradable insights faster than any human workflow.