Selected work

Project	What it shows	Stack
llm‑regime‑allocator	LLMs call macro regimes from point-in-time data to set ETF weights. Sharpe 0.83 vs 0.66 for 60/40 (not significant), audited for look-ahead memory.
multi‑asset‑portfolio‑lab	Seven allocation models (MV, Black–Litterman, ERC, HRP…) with shrinkage and vol targeting on 13 ETFs. None beats equal weight on Sharpe after multiple-testing correction.
ai‑pairs‑trading	OU-based pair selection on LatAm ADRs lifts Sharpe from ≈0 to 0.81; an earnings gate takes it to ~1.05. The news-sentiment overlay fails, and the repo shows why.
filings‑qa‑rag‑demo	Cited RAG over SEC filings: every claim traceable to a retrieved chunk, citations validated, OpenAI or Anthropic generation.
next‑gen‑agentic‑finance	Agentic research patterns in ten steps, ending in an evidence → interpretation → open-gaps handoff for human review. Read-only tools, no execution.
the‑ai‑engineer	Four capstones: gradient descent and backprop from scratch, a transformer measuring FOMC textual surprise, and an auditable MCP agent.