I build tools that keep the sensitive decision explicit.

I am a solo founder in Houston, building Atarla AI full-time: AI infrastructure you own.

I publish what I find out along the way, with the method and the sources attached.

Email meGitHubAll work

Portrait of Kevin Trinh.

Selected work

  1. Pixel-art thumbnail for Gaze: a browser window with a watchful eye

    Gaze

    Live2026–

    A command-line tool that drives the browser you are already signed in to, and asks before anything on the page changes.

    100/100 across 12 levels, including a prompt-injection level

    • CLI reads are free, writes stop for approval
    • MCP server the same gate, for agents
  2. Pixel-art thumbnail for Dispatch: one route branching into three

    Dispatch

    Live2026–

    A CLI that routes a task to the right AI provider CLI by data-sensitivity tier, cheapest eligible provider first, with verification before it trusts an answer.

    • Tiered routing explicit --tier, cheapest eligible provider first
    • Cascade with verification escalates when an answer does not look like an answer
    • Decomposition compound tasks split and routed at the same tier
  3. Pixel-art thumbnail for Inkboard: a chalkboard with hand-drawn strokes

    Inkboard

    Prototype2026

    A teaching-board recorder that keeps handwriting as data instead of baking it into video.

Writing

  1. Cameras, cars, DNA tests, phone records, password managers and AI chats all show the same pattern: when someone else holds the data, their systems become part of your risk.

Research

  1. Pixel-art thumbnail: an hourglass with pixels falling through

    Next up

    The next research note is being written. It ships with its method, its limits, and its sources, like the rest.

  2. Pixel-art thumbnail: a padlock with a broken chain link

    Public incidents show that data risk extends past the company you chose, to its credentials, employees, contractors, dependencies, connected devices, and legal obligations.

Elsewhere