career-ops: AI job search agent that scores listings, tailors CVs, and drafts answers

What they did

They used career-ops, the open-source AI job search agent, as their primary filter during a real search. Instead of blasting CVs into silence, they pasted job listings into the agent running on their own machine. For each listing the agent checked whether the posting was still open and whether it actually fit their profile, producing a scored evaluation report. Listings that failed the fit test were marked 'do not apply' and counted in red on the dashboard, so the noise was visible rather than silently consuming time. For the listings that survived, the agent tailored the CV to the specific role and drafted answers to application questions. The human stayed in the loop at the final step: they reviewed the tailored materials and pressed Submit themselves. Over the course of the search this produced a funnel of 740 listings evaluated, 68 applications sent, 12 interviews, and 1 offer — the author's own run, which they then open-sourced. The workflow is deliberately local-first: the agent runs on the candidate's machine rather than shipping personal career data to a third-party service, and the dashboard gives a single view of scored listings and rejected ones.

How it went

The documented run ended in a hire: 740 listings evaluated, 68 applied, 12 interviews, 1 offer. The project has since grown to 700+ contributors, a public HIRED wall of stories, press coverage from WIRED and Business Insider, and a Vercel OSS Program slot. Polish is high for an open-source agent — multilingual READMEs across ~20 languages, a demo GIF, and a dashboard UI. Limits: the funnel numbers are one person's search, not a benchmark, and the agent stops short of submitting for you.

Worth knowing

Results are self-reported from the author's own search and the public HIRED stories; there is no controlled comparison against a manual search. The agent evaluates and drafts but does not submit applications — you still review and press Submit, so quality depends on your own judgment of the tailored CV and answers. Fit scoring is only as good as the profile you give it, and stale or misleading postings can still slip through the 'still open' check. Setup detail (model/provider configuration, where data is stored locally) is not covered in the README excerpt and needs checking in the repo docs before you trust it with personal career data.

Wiring

Order of operations: install career-ops per the repo docs, then configure the model provider and confirm the agent runs locally. Next, build the candidate profile — the base CV and preferences — since every downstream score and tailoring step reads from it. Then the per-listing loop: paste a listing, the agent fetches and checks whether it is still open, scores fit against the profile, and writes an evaluation report. Listings below threshold get marked 'do not apply' and surface as a red count on the dashboard; listings above threshold go to tailoring, where the agent rewrites the CV for that role and drafts application answers. The human review gate sits between drafting and submission — the agent never presses Submit. The dashboard aggregates scored listings, rejected ones, and evaluation reports into one view. Keep the profile and generated artifacts in version control so you can diff tailored CVs against the base and audit what the agent changed.

Try it yourself

Set up career-ops in this repo and run it on a single job listing first. Steps: (1) read the README and docs to find the install and configuration steps, including which model provider it expects and where it stores data locally; (2) create my candidate profile from my existing CV/resume in this repo, keeping it factual — do not invent experience; (3) paste one real job listing URL and have the agent check whether the posting is still open and produce a scored fit evaluation with reasoning; (4) if it fits, generate a tailored CV and draft answers to the application questions, and show me a diff against my base CV so I can see exactly what changed; (5) do not submit anything — stop and hand it to me for review. Report the score, the reasoning, and any red flags about the posting.