I build AI products people actually use, and quant research that survives my own audits.

AI image workflow for a Shanghai commercial interior-design firm whose render pipeline was fully outsourced. Designers upload white-model renders and style references; the tool runs dual-engine generation, structure-preserving style transfer, high-fidelity retouch, and walkthrough video in batch — with per-call cost accounting and role-based accounts, so the owner finally sees where the money goes.

Hours of footage, minutes of signal. A local-first pipeline: quantized ONNX CLIP extracts 1 fps semantic embeddings, segments the timeline into visual states, and ranks segments by rarity, motion, and scene-change velocity — a 3-hour asset processed in ~12 minutes on CPU at zero cloud cost. Scored against a single-blind human annotation protocol: ranking AUC 0.82 → 0.93, precision@5 = 1.00, with the small-sample overfitting risk documented rather than hidden.

Cross-sectional factor research over 35 months of hourly data across 55 crypto perpetual markets: 14 pre-registered candidate factors, pass/fail rules written before any test, a fixed out-of-sample look budget. Four factors survived. Unleveraged Sharpe 1.34 full-period, 1.18 out-of-sample through a −22% market drawdown; Monte Carlo block-bootstrap ruin probability 0.3%.

The part I'm proudest of is the audit trail: I found same-period leakage in my own regime filter (sleeve Sharpe 6.75 → 2.43), retracted the result, and rebuilt selection leak-free. Every strategy now passes a 10-gate mechanical validation harness — magnitude, concentration, leverage, deflated Sharpe, lag, capacity, universe drift, Monte Carlo — first calibrated by feeding it deliberately leaky fake strategies to prove it catches them.

Route-based deep-culture travel companion: an AI narrator for the history under your feet, stop by stop, designed with a magazine editor's eye rather than a listicle's. Alongside it, a Socratic reading companion for serious nonfiction — one book, locked in, question by question.

• Lead-trader tracker — 60-second polling of a public Binance portfolio; open/add/reduce/close events pushed to Telegram in real time.
• Stock-perp radar — hourly cross-exchange scan of 250+ markets (Binance × Hyperliquid) for new listings, basis dislocations, and funding anomalies.
• Closing-line-value research — falsified six hypotheses over 249 matches before isolating a +3.5–6% CLV edge, consistent across six seasons and seven of eight leagues.