JobSieve · Augustus Forge

Why

I built it for my own job search. The first version was the obvious one: embed the resume, embed each vacancy, sort by similarity. It kept burying jobs I wanted and surfacing ones I would never take: right on paper, wrong on pay, format or plain honesty.

What came out

Two stages. First, does it look like my experience? Each section and project of the resume is matched on its own, so a data pipeline vacancy meets the pipeline project, not an average of everything. Second, would I actually take it? My own terms, with every hit written on the card and the reason next to it.

It learns from fit and no-fit decisions, follows what happens after I apply, and measures all of it on real decisions.

What broke

On my real search, 402 vacancies: inside a run the top ten went from about seven real fits to eight. Across runs it’s a coin flip.
Fourteen hand labels prove nothing yet. The number starts to mean something at fifty or more.
The evaluation caught one of my own rules boosting the wrong jobs.

The path

What I posted on X about it, oldest first.

16 Oct 2026Result

I built a tool for my own job search. Today it's open source.

The first version was the obvious one: embed the resume, embed each vacancy, sort by similarity.
It kept burying jobs I wanted and surfacing ones I'd never take: perfect on paper, wrong on pay, format or plain honesty.

So JobSieve ranks in two stages:
1. Does it look like my experience? Each section and project of the resume is matched on its own, so a data pipeline vacancy meets the pipeline project, not an average of everything.
2. Would I actually take it? My own terms: deal-breakers, salary floor, red lines. Every hit is written on the card, with the reason.

Then I measured it on my real search. 402 vacancies.
Inside a run it works: the top 10 went from about 7 real fits to 8.
Across runs it's a coin flip. And 14 hand labels prove nothing yet.
Best part: the evaluation caught one of my own rules boosting the wrong jobs.

All of it, the bad numbers included, is in the README.

Runs locally. No API key, no scraping, no auto-apply.
Done. Free, MIT. Use it or fork it.