About
I'm George, a software engineer in Boston, MA. I'm curious more than anything. I find something interesting, go deep on it, and usually end up building something small around it.
Lately that's been AI agents. ShallowSWE benchmarks coding agents on routine work, SandFS is a filesystem sandbox, MCPX turns MCP servers into shell commands. Before that it was spatial computing, Vim, and note-taking systems, which is what a lot of the older writing here is about.
I like the building part, and thinking about hard problems. That's most of what this site is.
By day I'm a Senior Software Engineer at AWS, on AWS Transform. Before that: AWS Transfer Family, Amazon Ads, and Yahoo. And before that, ten Windows and Windows Phone apps built with one other person in undergrad, past two million downloads.
Away from a keyboard, or at least a computer: mechanical keyboards, photos of sunsets, and the heavy bag.
If any of this is your kind of thing, don't hesitate to say hi. I always like hearing from people.
Start here
- Kimi K3 Reaches the Coding Frontier. Here's the Catch What attempt-level benchmark data says once you price it.
- Three Problems the AI Industry Is Glossing Over The human bottleneck, the quality ceiling, and the cost.
- The Eddy and the River Entropy, order, and where AI sits in the current.
Education
- 2018
University of Southern California M.S. Computer Science, Data Science specialization
- 2016
University of Patras Diploma, Electrical and Computer Engineering
On attribution
Where a piece was written with a model, the byline names it. EOO-004 through EOO-006 were written with Claude; EOO-007 with GPT-5.6 Sol. Anything earlier predates this and is mine alone.
The thinking, the claims, and the mistakes are mine either way. The byline is there so you know what you're reading.