A small, deliberate body of work at the seam of quantitative finance, machine learning, and physical engineering.
Quantifies geopolitical investment risk across 56 countries. World Bank macro data and live headlines, scored 0–1 by a rule-bound LLM, rendered on an interactive world map with a plain-English explanation for every country.
View case study →A Rust implementation of Combinatorial Purged Cross-Validation (CPCV) from López de Prado's Advances in Financial Machine Learning. It generates many combinatorial train/test splits with purging and embargoing to strip look-ahead bias from overlapping labels, producing a distribution of out-of-sample performance instead of a single fragile backtest — and an estimate of how likely a strategy is overfit.
A custom agent harness that turns a small, locally-run 4B model into a capable reasoner — cycling it through brainstorming, sandboxed code execution, self-critique, and synthesis on a time budget. Trades inference-time compute for accuracy on local hardware, and doubles as a testbed for reward-hacking research.