Job search

Running

Runs on Brain →

Finds openings across hundreds of employer job boards, screens each one against my written criteria, and builds a checked one-page résumé and cover letter for every pass. I submit every application myself.

  • 287employer job boards
  • 13public job boards
  • 5career lanes, rotated
  • 870+automated tests

HOW IT WORKS

Architecture

Code does the deciding, a local model does the reading, and nothing reaches my résumé unless it already exists in my own records.

Interactive diagram. Select a card to highlight what it feeds and what feeds it; press Escape to clear.

Who does the work:

DASHBOARD

The dashboard

Shown with demo data. The real one runs privately on my Mac.

Still
6 second walkthrough

UNDER THE HOOD

What it actually does

  • 287 employer boards, read directly

    Workday, Greenhouse, Ashby, Lever and more. Every listing is cached for the day, with time limits per source.

  • Fails loudly

    If a board that used to return jobs suddenly returns none, the run stops with an error instead of carrying on quietly.

  • No quote, no fact

    Anything the model reports is downgraded to unknown unless its quote appears word for word in the posting.

  • The model never decides

    Its output has no verdict field at all. Verdicts come only from code.

  • No numbers in the code

    Pay, hours and experience limits are read from my criteria file on every run. A missing limit stops the batch.

  • A guess can never reject

    An inferred field cannot reject a job, and conflicting findings go to incomplete rather than a coin flip.

  • Five lanes, rotated

    A fixed nine-slot cycle spreads the search across marketing, venture capital, consulting, finance and product, never leaning on one lane twice in a row.

  • Reader accuracy, measured

    Field accuracy went from 42% to 96% across four measured rounds, with every verdict correct on the last one.

  • Résumé lines cannot be invented

    The model returns line IDs only. Code prints my real lines, and a verifier rejects anything else.

  • Keyword matching with guardrails

    One model rewords a line toward the posting. Code blocks new numbers or tools, and a second model checks it against my full record.

  • Cover letters inside a fixed shape

    Four paragraphs, checked for invented numbers, banned verbs and length, then revised over several rounds.

  • Questions wait for me

    New application questions are never answered by the local model. Saved answers are reused only on an exact match.

HOW I DESIGNED IT TO WORK

Automated search, human hands

Three lines from the rules I set for the search.

“For every job it finds and it's a pass, the résumé should be automatically generated…”

On the dashboard

“To confirm accuracy with my work experience, it will do a second pass.”

On tailoring

“The apply side stays manual.”

On submitting

WHAT'S NEXT

Next, in order

  1. Use the rebuilt dashboard daily and list what still differs from the design.

  2. Let my own browser send postings from company sites that block programs, without faking a browser or beating a bot check.

  3. Clear out old résumé files once every active job is rebuilt.