Blog

Writing

Thoughts on building with AI, running infrastructure, and the messy reality of making things work.

Infrastructure/April 2026/6 min read

Why I Swapped a 122B Model for a 35B One — and Got Faster

Counterintuitive move: trading a dense 122-billion-parameter model for a 35-billion-parameter mixture-of-experts, and having the site feel noticeably snappier afterward. Active parameters matter more than total parameters.

Infrastructure/March 2026/8 min read

Running Production LLMs on a DGX Spark

What it actually takes to run large language models locally. Hardware choices, memory architecture, inference optimization, and why 128GB of unified memory changes the game.

Product/February 2026/6 min read

Why I Built a YNAB Alternative

Envelope budgeting is a solved problem. So why did I spend months building a new app? Because the existing solutions either cost too much or do too little.

AI/March 2026/10 min read

RAG Pipelines That Actually Work

Everyone is building RAG. Most of it is bad. Here is what I learned building a retrieval pipeline that handles real-world government data.