Building trading infrastructure where every statistic is honest by construction.
Flagship — multi-asset-backtester
🔴 Live demo → multi-asset-backtester.vercel.app (free-tier backend sleeps when idle — first load takes ~30–60s, then it's fast)
Multi-asset backtesting + multiplayer paper trading, deployed end-to-end:
- No-lookahead engine —
position = signal.shift(1)structurally, for user-submitted code too - Honest statistics — Deflated Sharpe reported against the actual number of trials attempted, purged + embargoed walk-forward CV, triple-barrier meta-labeling, VaR/ES with Cornish-Fisher tails, factor attribution before anything is called alpha
- Market frictions — commission, spread, slippage, and square-root market impact in backtests; Reg-T-style short margin and 2× leverage on a row-locked, Decimal-exact shared ledger in paper trading
- Bring-your-own-Python sandbox — AST allowlist (no imports/dunders/exec) with line-numbered rejections at submission time
- FastAPI · PostgreSQL · Next.js · WebSockets · 181 tests · CI green · $0/month single-process deploy
Now building — itch-lob-engine
NASDAQ TotalView-ITCH 5.0 order-book reconstruction in C++20: exchange binary protocol parsing, order-id book, Catch2 tests, Linux + macOS CI. Project rule: no performance number ships unless a committed, reproducible benchmark produced it.
Purged walk-forward CV with embargo ≥ label horizon · Deflated Sharpe against counted trials · transaction costs including square-root impact · factor attribution to separate alpha from exposure · negative results reported, not discarded.
Python (FastAPI, pandas, scikit-learn, XGBoost, PyTorch) · C++20 · TypeScript (Next.js) · PostgreSQL · Docker · GitHub Actions