TRULY ENGAGING
Finding the Real Capacity of a Die-Cutting Machine
Management wanted to know the daily capacity of two industrial die cutters. A year of machine sensor data showed that only 14% of available time was spent cutting, and that the constraint was changeover frequency rather than machine speed.
PERSONAL PROJECT
Testing Whether Earnings Volatility Can Be Traded
I built a research pipeline to test whether filtering earnings trades by company quality improves returns. The best run gains 79%, and almost all of that comes from a single year and a handful of outlier trades. Here is how I know, and why I reported that instead of the headline number.
DATA ENGINEERING
Building a Point-in-Time Stock Database
Every quant project I ran used to pull and clean its own data. I replaced that with one platform. A resumable crawler collected 1.9 million raw files, and validated loaders turned them into a 68 GB PostgreSQL warehouse of 8,700+ U.S. stocks back to 1970, where every number is dated to when it became public. New research now starts with a SQL query.
INFRASTRUCTURE
Building the Server My Research Runs On
One self-hosted Linux server runs every project I build. Tailscale keeps it private with a single login-protected public route, an Obsidian vault on it commits and syncs itself, and an MCP bridge lets Claude plan the work while Cursor writes and commits the code.
LOCAL AI
Running AI Coding Models Locally
Can a free model on my own CPU do real coding work? I built an agent pipeline around one and benchmarked three models. More CPU cores made them slower, reading long prompts was the real bottleneck, and turning reasoning off was the difference between a correct result and no code at all.