I build AI systems, and the checks that keep them honest.

I am building Kalvi to teach what you want to learn so it lasts. It tests what you can do with and without tools, then lets you prove it. I have written 4 books on Indian philosophy and attention.

Start here A check that cannot look Every serious bug in my personal agent was a check that reported clean because it could not look. Nine cases, why the shape recurs, and what I do now. 7 min read

Latest writing

  • 7 min read

    Voice typing on Wayland with a pair of EarPods

    Voice typing on Omarchy with Voxtype and wired EarPods: per-device Hyprland binds, a click dispatcher for the remote, and saying submit to press Return.

  • 9 min read

    The colors a design system does not have

    Porting this site's design system to an Omarchy desktop theme: the colors that mapped, the slots that had no source, and the muted key I nearly got wrong.

  • 9 min read

    Three ways coding agents draw a status bar

    How Claude Code, Antigravity, Codex and pi each let you customize the status bar, how three of them share one renderer, and why Codex cannot join.

  • 4 min read

    A longer prompt is not a safer prompt

    Why adding sentences to a prompt can make a model worse, what Chroma's context-rot study measured, and an editing pass that shows which lines carry weight.

All writing RSS 14 essays and notes, newest first