Selected work

9-agent crypto quant pipeline. Best signal hit Rank IC 0.0275, but after walk-forward validation and costs it reports alpha_verified=false.

Machine forecast-disagreement study on 397K JKP stock-months. Accounting-only vs market-only Random Forests drove a 1.19%/mo portfolio spread. Won Project Champion.

URECA paper testing Perplexity-built S&P 500 portfolios. The high-risk book returned 31.5% over three months, but the edge came from sector tilts, not stock-picking.

Pre-submission healthcare claim denial-risk engine trained on real CMS CERT audit data (ROC-AUC 0.745, PR-AUC 0.295 vs 0.141 base rate). CARC/RARC crosswalk, expected-recovery-$ queue, model-card page, 76 tests.

Macro-risk dashboard fusing live keyless ETL (World Bank, GDACS, UNHCR) across 60 countries into an INFORM-aligned, geometric-mean 3-dimension risk score with a real historical trend store.

Prediction-market agent where the LLM only labels headline sentiment and Python computes the Bayesian edge versus market price. Public track record joined to real resolutions, admin-gated controls, 77 tests. Paper trading; deploy pending.

Agent payments and escrow app built at ETH Global. On-chain x402 receipt verification and a pre-signing spend-policy engine (per-request + daily USD caps) now gate every payment route. 154 vitest + 31 forge tests, testnet-only.

Ingests ARK ETF holdings and SEC 13F filings, flags NEW/EXIT/INCREASE/DECREASE position changes, and warehouses them to Parquet.

Polymarket signal oracle that sells its Bayesian edge for $0.01 USDC per call over x402 on Base, now with a public track record joined to real market resolutions. Alphanet Core, its diverged equities-data twin, verifies x402 settlement on-chain and books its own unit-economics P&L.

Real-time multiplayer trading sim. A Rust/SpacetimeDB backend ticks a GBM market with volatility regimes and limit/stop orders every 2s, server-authoritative; React Three Fiber renders a walkable 3D city. 48 cargo tests.

Risk terminal for athlete finances, now backed by live NBA data (BALLDONTLIE) and a k-NN comparables engine. An honest backtest shows the collapse heuristic beats base rate ~2x but doesn't clearly outperform a simple age-31 rule — reported as-is. 63 tests.

Enriches 801 hospital equipment rows with per-vendor serial decoders, resolving 73% locally and calling OpenAI only on the gaps.

Crypto forecasting study on real Coinbase/Kraken OHLCV. No model beats a persistence baseline on next-day returns; the durable signal is regime clustering (DOGE silhouette 0.625). 23 tests.

NLP fraud classifier on 17,880 EMSCAD job postings. 0.80 precision on the fraud class and 0.66 PR-AUC against a 0.048 base rate. 19 tests.

RAG chatbot over YouTube transcripts. Hybrid BM25+dense retrieval with cross-encoder reranking lifted Recall@1 to 0.673 and MRR to 0.788 on a 55-query eval set (from 0.509/0.684 dense-only), with a CI gate to prevent regressions. 44 tests.

Lab pipeline that turns HR .docx files into structured CSV and scrapes public contact emails across a three-notebook workflow.