Thoughts on code, creativity, and the chaos in between.
How I built a semantic search engine for historical fashion collections using Claude, CLIP, and vector embeddings — and what MindCap taught me about input quality.
The hardest part of working with AI isn't the code — it's the context. Here's the documentation system that made long-term AI collaboration actually work.
Full-stack refactor complete: replaced a 15-category system with five behavioral intents, deleted 1,212 lines, and caught a silent bug hiding in the pattern detector.
Two phases into replacing MindCap's categorization system with intent detection. Simpler, broader, and half the code.
I ran keyword extraction against 56,000+ real browsing records and found the system only understood programmers. Here's how I redesigned it.
Rethinking topic categorization by prioritizing content signals over domain signals, plus URL garbage filtering.
Development diary on rabbit hole tracking, keyword extraction, and the cognitive overhead of building alone.
Escaping Intermediate Hell, rebuilding in React, and rediscovering my voice with the help of AI.
The biggest, most ambitious project I've ever built—and an honest look at the difficulties of inferring attention from observable behavior.
Transforming raw browsing sessions into meaningful insights about what topics capture your attention and how that interest evolves.
The story of building a browser extension to understand where your attention actually goes online, and what it's like to develop with Claude as a thought partner.
Building a Python tool to crawl 56,000+ URLs from my Firefox history and discover what I've actually been doing on the internet.