ShortList

The problem

The job boards are a firehose, and the good jobs hide behind the wrong titles.

Seven boards produce tens of thousands of postings a month. Most are impossible on their face — the wrong country, a salary under the floor, a management job wearing an engineer's title. The rest need reading, and the best ones are the ones a title search would never find: an "Accounting Automation Engineer" posting is a great job for an analyst who builds, and no keyword list knows that.

The usual answer is to skim, get tired, and apply to whatever was on top. The tool exists so that the reading happens every morning, on everything, with a reason and a doubt written beside each score.

The solution

Code removes the impossible. The model judges what is left. A person has the last word.

Every posting from seven boards lands in one local database. Deterministic rules hide what could never be taken, and every hidden posting records which rule hid it — switch a rule off on the page and its postings come straight back, no refetch, no rescore. The survivors are then read by a model and scored on two axes: fit (do the skills match what they asked for?) and thrive (would the work actually succeed?). The top slice is re-read twice more and the median kept, so one wild reading can't promote or bury a job. Everything at or above a fixed bar is the shortlist.

Beside the list sits a chatbot that talks to the model about the posting on screen. Its answers are grounded in the database — the score, the stated doubt, the rule that fired — and it is told to say "I don't know" rather than infer. Every message is kept, because the observations worth having are the ones you have while looking at a posting, and those are the ones lost by the time you switch windows.

Then the part that doesn't move: the tool never applies. It saves, dismisses, and — on approval — reorders the résumé for a posting from one source of truth, never adding a fact that isn't there. If a posting asks for something the profile doesn't have, the line is left out. There are no cover letters, and a test reads the source tree to make sure no code path defaults to one. Every application that goes out is recorded — link, title, date, the exact résumé sent — which is also the state's work-search record.

Status

Private, and in daily use.

Built for my own search and run every morning. The counts on the anatomy page come straight out of the database, dated. It stays private — the data is one person's job hunt — so instead of a launch link there is a forty-second recording of it running on a scrubbed copy of the data, and three stills from the same run.