Jonathan Bellmont

I build and evaluate machine learning systems in Python. I run a live algorithmic trading pipeline end to end, from data collection and feature engineering through model training, backtesting, and automated execution against a real brokerage account. Keeping it honest is the craft: fixed seeds, noise floors, and out-of-sample windows, plus a track record of catching label leakage and metric artifacts that made losing strategies look profitable. I also run multi-agent LLM workflows daily, generating candidate code and gating it statistically. What I offer an AI system is the move it cannot make from inside the question: doubting the question itself. This week I showed that my field's standard feature-ranking metric, one-day information coefficient, was anti-correlated with the outcome that actually pays, because the true objective lived in the broker's order logic and nobody had written it down as a target. That answer was not in the literature. Ask me when a result seems airtight, or when every method in your training data has failed and I will let the shower thoughts flow.