How Fixing an OOM Error Quietly Broke My Model
I capped my vocabulary to survive a Kaggle OOM. The loss dropped, the accuracy looked fine — and the model had learned nothing but token frequencies. A debugging story.
Writing
Deep dives on distributed systems, ML infrastructure, and the craft of software.
Setting Up This Blog
How this site is built: Astro, Markdown content collections, GitHub, and Cloudflare Pages — and how to write your next post.