· 2 min
writing
i write to find out whether an idea survives contact with a real system, so most of these start from something i traced, broke or shipped rather than something i read. the register is deliberately plain: name the failure, show the mechanism, then say what the mechanism does not prove.
long-form on x
everything long-form goes to x first. the twelve on the writing wall are the ones that still hold up on a second reading — each one entered through a run of posts, and each links back to the thread it came out of.
- 4,264 — posts published since joining in february 2022
- 12 — long-form pieces selected for the wall, each checked against the live page
the contribution trail
the posts i trust most are the ones with a pull request under them, because the artefact is checkable by someone who has never met me. they are not all on the wall — several sit under a bare link card with no headline a page can use — but they are the part of the record i would point at first.
- 10 — posts carrying a pull-request link into a repository i do not own
- 4 — upstream projects those went to — mem0, superset-sh/superset, mcp-use and deepeval
selected pieces
- I Audited Mem0. The Hardest Problems Lived Outside Retrieval (opens in a new tab) · 18 august 2026
- A Second Prompt Is Not an Independent Code Review (opens in a new tab) · 5 september 2026
- Your AI Agent Has Five Retry Loops. One Failure Can Become 243 Calls (opens in a new tab) · 2 august 2026
if any of it is worth arguing with, hit me up.