Context
This is a job-search agent I wrote as the sole author for a designer's job hunt. It is in use, and it runs locally.
The problem
A job hunt needs discovery, fit scoring and applications at volume. Done carelessly, that produces fabricated claims and spammy behaviour, which hurts the applicant.
What I did
I built the pipeline: discovery, fit scoring, resolving each posting to its real application system, referral lookup, a tailored application package, a policy gate, then submit or send to review.
I wrote the resolver for four applicant tracking systems: Greenhouse, Lever, Ashby and Workable.
Generated text goes through a numeric fact-check. Numbers in a draft are checked, and a failed check falls back to a template.
I added score calibration using Wilson intervals, ghost-job detection and liveness checks, rate limits and a full submission audit record.
Key decision
The key decision was a strict policy gate between generation and submission, with fact-checks that fall back to templates. The cost is fewer fully automatic submissions. The gain is that nothing invented gets sent.
I judged that trade-off worth it because a wrong claim on an application is hard to take back.
Result
It resolves postings across 4 ATS vendors and is in use. I do not report applications sent, interviews or offers, since that is the applicant's personal data.
What I learned
I would again keep the checks in code, not in the prompt. I would add outcome tracking early, with the applicant's consent, so calibration can be judged against real responses.
Stack
- Next.js
- TypeScript
- Prisma
- PostgreSQL
- Playwright
- Claude API
- MCP
