2. jobFinder

Each day the system brings in new ads automatically as soon as my computer is connected to power. Closed ads are removed and ads that pass the preliminary filters get passed to the model for rating. Every list is sorted with the best job for me first. When I click on a job, the two ratings and the model’s reasons appear next to the list in short form, together with a link to the original ad and one button “Write brief”, which starts an application. Most days I only look at what is new, save a few jobs and start an application for one or two. About once a week I go through everything more slowly, including the jobs the model rejected, in case it has missed something that I would find interesting.

The job title is no help in sorting them, because the same work is often advertised under different titles. For example “Systemförvaltare”, “Business Systems Analyst” and “Analytics Engineer” might get flagged by the model for multiple lanes, which is also presented when opening the job card. See below for an example.

Most job boards rank ads by how well they match the keywords you use while searching. What I needed was a ranking by how well each job fits me as a whole. So the model reads every ad next to a written description of me and gives it two ratings from 0 to 10. Want is how much I would want the job. Reach is how likely I am to be hired.

The description states my limits as plainly as my strengths, because a description that oversells me would put jobs I will not get at the top. It also says what I want from a job, such as owning a process from start to finish, senior people to learn from and a slight preference for smaller organisations. The model can read those things in an ad, where a keyword filter cannot.

At first the model mixed the two ratings up. A job far above my experience came back as a job I would not want, and telling the model that the two ratings were separate did not help. What worked was asking a different question. The instructions now say: “Imagine he has already been offered this job and has already accepted it. He starts on Monday. How glad is he, on his first morning?”

Every list in the dashboard is sorted by one score made from the two ratings, and that score is not an average. An average would put a job with very high want and low reach next to a job as a sales assistant in a shop, because the two come out almost the same, even though I would take the first any day and never the second. So want counts the most, and a low reach only lowers the score, by up to about half.

Every ad the model reads costs money, about a third of a cent each, so before anything was built I set a limit of 40 USD a month and designed the system to stay well below it. The money is paid in advance with automatic top ups turned off, so the system can never spend more than what is already there.

The first saving is to skip the ads that cannot fit before paying for them. Every ad is stored, since storing costs nothing, but only ads from the job categories that could fit me are sent to the model. I went through all 159 of Arbetsförmedlingen’s job categories and marked 96 of them to be skipped, so about 40 percent of the ads reach the model. The categories are sometimes wrong, and a trainee AI consultant ad that was close to a perfect match had been filed under telephone sales. So an ad from a skipped category is let through anyway when its headline names the kind of work I am looking for. To test the skipping, 100 skipped ads were rated anyway, and none of them scored higher than 3 out of 10 on want.

The second saving is a cache. The same instructions and the same description of me are sent with every ad, and together they are longer than the ad itself. Because they are stored and reused, they are billed at a tenth of the normal price.

Third, nothing was rated in full until a smaller test had been approved: first 17 ads that I had rated by hand, then 400, and only then all 6,571 open ads that got past the skipping, on 12 September 2026, for 20.43 USD. And no ad is paid for twice. Only ads the model has never read are sent to it, however many arrive in a day. Between 14 and 17 September 2026, rating a day’s new ads cost between 0.13 and 1.62 USD.

I will be running this every day until I get a job offer, so it has to take in new ads on its own and send only the new ones to the model. One command runs all the steps in order: it brings in the new ads, marks the closed ones, skips what cannot fit, has the model rate what is new and rebuilds the ranking. Every run ends with a short report of what arrived, what closed and what it cost, compared with the run before.

From 14 September 2026 the run started at 06:00. My laptop is almost never awake and plugged in at that hour, so most of the runs that finished took six hours or more, and two broke off after the first step. Since 22 September 2026 it starts instead the first time each day the laptop is plugged in, and keeps the laptop awake until it is done. Nothing was lost on the bad days, because every run picks up from where the last finished run stopped.

Since 16 September 2026 it also reads ads from other Swedish job sites, through a second open service called JobAd Links. It was said to add around 30 percent more ads. When I counted, it was 4 percent, but they were ads I would not otherwise see, such as startups on thehub.io and junior roles on studentjob.se. Those sites only share the first few lines of each ad, and I did not want the model rating a snippet, so the full ad is read once from the site itself. An ad that is also on the main API is dropped before it costs anything. On 17 September the two jobs at the top of the whole list both came from thehub.io. Future development on this software will add APIs from other countries as well.

While designing the front end UI I wrote down what the dashboard needed: a switch between the best jobs overall and the best in each lane, a star to save a job, the last day to apply, and how many ads had come and gone since yesterday. I also wanted it to live as files in my own folder and run on my own laptop, not as something published online. I wanted the dashboard to be very simple and usable, without any distractions or any unnecessary features.

The first step was a plain sketch with real ads in it. The information on each card was right, but too much showed at once. All seven kinds of work were stacked on one page, and each list was filled with up to 20 jobs even when only a few of them were jobs I wanted, which made every list look fuller than it was.

First, I went through the wireframe build and changed what did not work. Then I went ahead and built the real dashboard in React, since this is the front end framework that I know the best. Usability is key, one of the areas that I focused heavily on over the past years, and wanted to focus on here as well. The main inspiration for this dashboard is a very stripped down version of Airbnb adapted to include only the functions that I needed. A clean and modern dashboard aesthetic.

During the build a serif typeface replaced the first one, the tile with the employer’s initials went, the header got the name jobFinder and a logo picked from four options, and the tall cards became compact ones, with each row only as tall as its longest card. Loading more jobs as I scrolled filled the large monitor with 60 cards at once, so the list now opens with ten and a button adds ten more in order to reduce cognitive load.

I also added a view of what is new since I last looked, a switch between Stockholm county and everywhere else, since any job outside Stockholm means moving, and a list of the jobs the model rejected, so that nothing it read disappears without my seeing it. Then I paused the design work, because I reached a place where the dashboard did what I needed.

The dashboard never sends applications in my name. When I click Write brief on a job, it makes a folder on my computer for that job with a short brief in it, holding the ad, the ratings and the model’s reasons, and then it stops.

I then open Claude Code in that folder and ask it for a CV and then for a letter. Both draw on a bank of about a hundred short claims about me. Each claim is something I would stand behind, together with the evidence for it and the question it is likely to raise in an interview. The bank also lists what must not be claimed, so that wrong claims cannot find their way into a CV.

The CV is put together from those claims, but the letter is not. Letters assembled from ready made parts had once given me two letters, to two different employers, that were the same letter with a few words swapped. So each letter is written new for one job, using the claims only for facts, and it draws on the claims the CV had no room for. It is held to a description of how I write, built from references of my own writing. I read and edit every CV and letter before it goes anywhere, since I consider AI writing sloppy at times.

Design and every decision: Oscar Ström. It is my own project, started on 4 September 2026. The code is written with Claude Code from my design.

Job ads: the open data of Arbetsförmedlingen through JobTech Dev, and other Swedish job sites through JobAd Links. Ratings: Claude Haiku 4.5.

Design process: Impeccable. Typeface: Newsreader. Icons: lucide. The CVs and letters are set with Typst.

The screenshots on this page were taken on 17 September 2026 from a copy of the data, with my saved jobs cleared. The ads are real and public. The terminal images are the real output of the commands, set as text.