Open to ML / LLM and recommendation-systems roles — staff or senior, ideally where agent reliability, evaluation, or retrieval is the hard part. Remote or Bay Area. The fastest path is a 20-minute call; I reply within a day.
01What I’m building
- Agent reliability infrastructure. Memory that survives compaction, evaluation that discriminates between agents rather than scoring them, and the loop machinery that lets a run last hours instead of minutes. Most of it ends up as a post with a working simulator in it — the start-here page has the current five.
- Trace visualization as a shared component. A D3 span-waterfall viewer deployed across several posts, so an ideal run and a doom loop can be compared as shapes rather than as log files. Built once, mounted anywhere, all client-side.
- Contributing upstream. Recently a fix to Prime Intellect’s prime-agent — an API asymmetry in its Python runtime where the delete call rejected the handle its own spawn call returned. Small, tested, and written up honestly in The Last Click, including the four permission walls it hit on the way to being publishable.
02What I’m thinking about
- Verification is the bottleneck, not generation. Models got good at producing plausible work faster than we got good at checking it. Nearly everything I write now circles this: how you tell a finished task from a claimed one, and what it costs to know the difference.
- The harness outlives the model. Products swap models constantly and remain themselves, which means the durable engineering — and the durable career — is in the harness rather than the weights.
- Permission boundaries as interface design. After watching an agent hit four walls trying to publish a patch, I’m increasingly convinced the interesting design question is not how much autonomy to grant but how much value a single human consent click can be made to convert.
03How I work
- Everything ships with something runnable. Each post carries an interactive artifact you can use in the browser — a simulator, a grader, a playground — because a claim about agent behaviour is worth less than a thing that behaves.
- The blog has a test suite. Every push runs a headless browser over all 23 posts and checks that each loads, that its widget actually boots, and that no link is broken. It fails the build otherwise, and it has caught three real regressions — including one I had shipped myself.
- Measured honestly. This site sets no cookies and runs no trackers; the analytics are cookieless and aggregate. I’d rather have smaller numbers I believe than bigger ones I bought with someone else’s privacy.
04Elsewhere
GitHub · LinkedIn · Google Scholar · RSS · sugeerth@gmail.com