About
I love my wife and our two sons.
I love computers and music.
I grew up in rural Sweden and now I'm back in rural Sweden, after a 13 year detour in London, Boston, NYC, and SF.
I work as a principal engineer at Cloudflare. In 2025, Cloudflare acquired the startup I co-founded, Replicate. Replicate is going strong and I'm still deeply involved.
Prior to Replicate, I was a machine learning engineer at Spotify, where I built the first deep learning stack, wrote some papers and shipped some products.
I have a PhD in machine learning for music source separation.
I'm a licensed fork lift driver. I like to grow potatoes and pick blueberries. I play accordion and guitar.
Assorted work
cctr
cctr is a test runner for command-line tools. Tests are defined as plain text corpus files that specify commands and their expected output.

plwr
plwr is a Playwright CLI for browser automation using CSS selectors. It's designed to be ergonomic for both humans and agents. Paired with cctr it makes for a powerful end-to-end testing tool.

leta
leta is a command-line LSP client for agents to navigate and understand code. It has fast symbol search, reference finding, call graph analysis, etc. Makes agentic search a lot more token efficient.

veta
veta is a command line tool and Cloudflare worker service for agents to take notes. Inspired by beads but without the bloat. Put veta ls or veta tags in your system prompt for persistent memory across sessions.

QQQLANG
QQQLANG is a programming language I wrote for my three-year-old where any string of characters is valid program, that generates an image.

greger.el
greger.el is an Emacs major mode for agentic coding. Every chat is a self-contained file in human-readable markdown-like format, every file edit is a new Git commit. I'm working on a new version with subagent support, but never MCP.

Replicate
Replicate is the thing I'm most proud of in my professional career. Also the hardest thing, by far. Ben saw the deep learning research I was doing at Spotify, and had the incredible foresight to bet on AI before almost anyone else. We created a new market segment and we defined the interfaces for model serving that have since become standard. We joined Cloudflare to build Replicate on the world's best network and developer primitives. The network is the AGI.

PhD thesis
I wrote my thesis "Musical source separation with deep learning and large-scale datasets" part-time between 2011-2019, back when we were painting the map of deep learning architectures. It was a fun time, you could publish papers just by applying deep learning to new domains and datasets.

This Is My Jam
This Is My Jam (2011-2015) was the best music site that ever existed. Founded by Matt and Han, we built a community of music nerds who shared their favorite songs. The level of quality was extremely high. Maybe we didn't have the mass appeal that the "market" wanted, but why does everything need mass appeal?

Surely there are phenomena not...
Surely there are phenomena not commercially viable nor interesting from other points of view.
Surely there are problems that can't be blamed on others nor make claims on their attention.
Surely there's a sorrow so felt it's disgusting.
Kristina Lugn, translated by Zach Maher