The open source app that ran my job hunt 1,008 times

The problem

The first version found jobs and wrote resumes. Then I used it every morning, and the gaps showed up one at a time.

My resume checker gave my real resume 100 out of 100. ResumeWorded gave the same file 73. Recruiter replies sat in my inbox while the Sheet still said Applied. Adding a company meant editing a config file.

I was the only user, so the spec was whatever had annoyed me the morning before.

What I built

JobPilot grew into a small product. A Cloud Run job polls 163 company boards and four aggregators. A Google Sheet is both the database and the dashboard. A resume judge drives a rewrite loop, an inbox watcher reads replies, outreach becomes Gmail drafts, and a console with a copilot sits on top. It is MIT licensed on GitHub, with a fork guide written to be handed to a coding agent.

Open the drawing to see every part. The live demo runs the real console on fictional data.

The sprint

v1.0 went public at 7:12 PM on June 10. The inbox watcher went in that same night and was done at 9:56 PM. By 11:10 the next morning it had a guard for one time passcode emails, after it alerted me twice that a company had replied when the company had only sent a login code.

June 11 brought the company watchlist and four more ATS types: Workday, SmartRecruiters, Workable and Recruitee. At 12:19 AM on June 12 the job got 2 GiB of memory, because the parallel fetch across 163 boards ran out at the default. The copilot shipped that afternoon and the US only location wall at 7:55 PM. Cold outreach landed at 4:41 AM on June 18.

The busiest day came a month later. On July 23 an audit of the Sheet turned into a rebuilt console, and July 24 went into knowledge graphs of my portfolio and GitHub to ground the copilot. The last change went in at 3:23 AM on July 25.

The hard parts

A judge I could check against a real one. The first checker only measured keyword coverage, which is why it loved everything. I rebuilt it as a replica of ResumeWorded’s rubric: about 40 checks, weighted impact 35, brevity 20, style 15, sections 15 and soft skills 15. Calibrated against the real thing, it scored my old resumes 76 to 86 where ResumeWorded said 73. Then every tailored resume goes through a loop:

# src/jobpilot/rewrite_loop.py
report = judge(tex, pdf, keywords)
if best is None or report["score"] > best[0]:
    best = (report["score"], tex, pdf, report)
issues = report["issues"]
if passes(report, min_score):
    break

The issues go back to Gemini, which may reorder and rephrase but never invent. Up to 10 tries, an early exit at 90 with 85 percent keyword coverage, and the best attempt wins. My real ResumeWorded score went from 73 to 82 after one pass.

The same job, five times. One role sat in the Jobs tab five times, and there were 23 duplicate groups in 426 rows. Exact duplicates were zero. The key was the problem: it hashed company, title and location, and boards list one role per city. The key is now company and title, and the Sheet’s old ids are recomputed from those two columns, so rows saved under the old key are recognized.

Failures that made no sound. The fast fetch never ran once. Its trigger passes arguments to the job, which needs a permission the plain invoker role lacks, so every attempt returned 403 to nobody. The console’s Tailor and Draft buttons failed the same way. One narrow role on the job fixed both, and the bug log tells every fork to check job status after any IAM change.

What I said no to

Guessing email addresses. Outreach uses a careers email the company published, found on its site, in the job post, or through a web search. No recipient means no draft.

A database migration. The July audit found 5,093 rows in the Jobs tab and 2,960 New rows older than 30 days. I kept the Sheet and added an Archive tab, a 14 day window and a fit 75 gate. Dedup reads both tabs, so nothing archived comes back.

Big tech portals. The seeded watchlist leaves out the largest companies with custom career sites built to block scrapers. The aggregators already cover them.

Where it landed

It ran my search every morning. The pipeline ran 1,008 times from June 10 to October 8: every 30 minutes through June 25, then twice each morning, at 7 and 9. By the July audit it was adding 22 to 58 fresh rows a day.

In September, its last full month, Google billed $11.64 for the project, and $7.86 of that was Gemini. My own copy has been off since October 8. The code stays public, and the demo runs the console on fictional data.

I never measured hours saved or which interviews came from it.

What I would keep: run the cheapest filter first, check any judge against a real grader before trusting it, and keep a person on the send button.