The Job Automation Index
A task-level index ranking 100 U.S. occupations by how exposed they are to AI automation.
Ranking jobs as a whole hides the real picture. Break a job into its actual tasks and the exposure to AI automation varies enormously within a single occupation.
Background
The idea started from Andrej Karpathy's point about verifiability: AI advances fastest on tasks where it can attempt something and easily check whether the result is correct. Instead of asking "will AI take this job," I wanted to ask a more precise question at the task level.
How it works
The index breaks 100 occupations down into five present-day tasks each, then scores how exposed every individual task is to AI automation, rather than scoring the job as a single blob. That task-level view surfaces which parts of a job are genuinely exposed and which are not, which a whole-job ranking would flatten out.
Technology
Built in Lovable across three phases: the leaderboard, the interaction and filtering layer, and the methodology and feedback pages. The backend runs on Supabase, including a feedback table protected with row-level security. Later builds moved the codebase to Codex. Built on the product specification drafted in Lovable.
How it was built
I built this spec-first, starting with a discovery build in Lovable to get a full working version fast and surface the gaps in my own spec: mobile layout issues, jargon that needed plain labels, and a schema I had guessed wrong (rank and id turned out to be different things, and several fields I assumed were numbers turned out to be formatted strings). Fixing those gaps in the spec rather than only in the code was the real work. Terminal agent builds followed once the spec had hardened.