Senior AI Engineer. Physics PhD in Quantum Chaos. Ex-Citadel Engineering. Writing & building at the edge of LLMs, data platforms, and indie software.
“First-principles from physics, gut-feel from years of debugging hundreds of billions of records.”
My work spans applying LLMs to improve data/ML pipeline reliability, using first-principles physics to plan petrochemical plant construction, developing hedging strategies for insurance products at a publicly traded company, and debugging data-quality issues inside hundreds of billions of records at a hedge fund.
Currently focused on orchestrating AI agents to build a production SaaS platform, and writing about what works when the tools change every three months.
If mental math is not about talent or random tricks, then the training question becomes simpler: train facts, patterns, estimation, and calm repetition.
People who are good at mental math are usually not doing the same arithmetic. They are doing easier arithmetic, because they see how to simplify the problem.
Most mental math advice treats arithmetic like stage magic: memorize a few tricks, think faster, impress your friends. That misses the point.
Practice everyday math in a fun way.
A platform with hundreds of everyday mental-math games — arithmetic, finance, sports, daily scenarios — built for both solo skill-building and real-time multiplayer. Typed input, no multiple choice, not LLM-cheatable.
Column-level lineage for SQL.
A Python library that extracts column-level lineage from SQL queries through static analysis. Understands which source columns feed each output column — without executing queries or connecting to databases.
End-to-end streaming & batch pipeline.
A production-style platform demonstrating modern data engineering practices: streaming + batch ingestion across SaaS sources (Shopify, Stripe, HubSpot), cross-source entity resolution, and a multi-layer medallion pipeline from Raw through Marts.
16-tile mahjong versus three AIs.
Single-player 16-tile Taiwanese mahjong against three AI opponents whose skill levels are independently adjustable — a full 16-hand match with dealer rotation, 連莊 continuation, and authentic tai (台) scoring. One TypeScript codebase runs everywhere: static web build and genuinely native iOS/Android via Expo, not a webview wrapper.
Interactive 3×3 solver, step by step.
An interactive 3×3 Rubik's Cube solver built on the Bevy engine. Enter the cube's colors or a scramble, let it validate the configuration, then watch a step-by-step animation solve it — via a human-friendly layer-by-layer method or a near-optimal two-phase (Kociemba) solver. Runs as a native desktop app and in the browser via WebAssembly.
Interactive 3D LEGO-style text generator.
Type any text and watch it construct itself, brick by brick, in an animated Three.js scene. The hero on this page uses the same engine.
The same idea, re-imagined in the terminal.
A Rust terminal app that transforms text into ASCII art using brick-like blocks and Unicode box-drawing characters, with drop physics and marquee scrolling.
Open to new collaborations, conversations about data systems, LLM infrastructure, or the kind of side projects that turn into real businesses. Reach out — I reply.