Blog
Notes on the craft of building.
Essays on machine learning, systems design, and the human side of shipping software.
- May 12, 20268 min read
The eval set is the product
Why your offline evals deserve the same care as your features — and how to grow them without burning out.
- Apr 02, 20266 min read
Cheap context windows are a tax on bad retrieval
Bigger context isn't a free upgrade. A note on what gets worse when you stop being picky.
- Feb 18, 202611 min read
A practical guide to small models in regulated domains
Latency, audit trails, and the joys of running a 3B model on a single GPU.
- Dec 04, 20255 min read
What I learned shipping fraud ML at 2B events/month
Three lessons from a year of streaming features, drift, and very angry analysts.
- Oct 20, 20254 min read
Notes on writing model cards people actually read
Stop with the markdown wall of tables. Start with a paragraph.