Notes & project deep-dives

A pipeline that catches denial risk before a claim ships. Gemma extracts structure from documents, XGBoost scores denial likelihood, a DuckDB knapsack query orders the review queue.

A prediction-market agent where the LLM labels news sentiment and Python computes a Bayesian posterior. Edge is the gap between that posterior and the market price. FastAPI backend, React client, paper trading.

A global crisis dashboard that pulls disparate risk indicators into one geospatial view. React and TypeScript front, FastAPI back, an interactive globe, and threshold alerts.

SHF-ML measures forecast disagreement between accounting-only and market-only Random Forests on the JKP dataset. Won Project Champion.

Co-authored URECA paper on Perplexity-built S&P 500 portfolios. The high-risk book hit a 7.545 Sharpe, but the returns trace to sector tilts, not stock-picking.

Privy-authorized agents pay $0.01 USDC on Base for Bayesian Polymarket signals. Usage-based machine commerce with no API keys or subscriptions.

Project CHF v2 reports alpha_verified=false. Why a pipeline built to reject weak alpha claims is worth more than one built to find them.

TF-IDF + VADER features on 17,880 job postings: 0.80 fraud-class precision and 0.66 PR-AUC against a 0.048 base rate. Why accuracy is the wrong headline for rare-event detection.