Swipe2Job
A self-hosted app that shows Jobindex and Jobnet jobs one at a time and learns from my swipes. A React and FastAPI PWA running on my homelab.
Jobindex felt kinda old school to me. Too much info everywhere. I wanted job hunting to be more fun, so I borrowed the way Tinder does it: you only have one job in front of you, and you say yes or no.
There was a second problem too. I only ever looked through a few categories. Most of the jobs in marketing don't spark any interest in me, so I'd skip that whole category. But a marketing job where you make content with video skills, like a videographer? That I'd want to see. Video is one of my hobbies. On Jobindex I'd never find it, because I'd never look there.
One job, yes or no
Swipe2Job scrapes every job in Region Midtjylland from jobindex.dk, which is about 4,000 of them. The first start runs a full backfill at one request every 2.5 seconds, so it takes around 15 minutes before the stack is complete. A scheduler keeps it up to date after that. Since October it also pulls jobs from Jobnet (more on that below).
Swipe right to save, left to reject. Every swipe can be undone, and the card flies back in. Tap a card and a sheet slides up with the full ad and a map of where the job is. Block words hide jobs I never want to see, and boost words push the ones I do.

The videographer job hiding in marketing
That category problem is what the ranking is for. Every job gets an embedding from bge-m3, running locally through Ollama, so the model sees what a job is about instead of which category it was filed under. My swipes are the training data.
Two scorers compete. One compares a new job to the jobs I've saved and rejected. The other is a logistic regression classifier that also knows the distance from home. The app tests both on my most recent swipes, which neither of them trained on, and keeps whichever does better.
Not every card comes from the ranking. Some are wildcards, so the stack doesn't turn into more of the same. Ranked cards get a match line naming the saved job they're closest to. And an About me text gives the model something to go on before I've swiped much.
1def score(self, X: np.ndarray, meta=None) -> np.ndarray:2 s = mean_topk(X @ self.pos.T, self.k_pos)3 if self.neg.shape[0]:4 s = s - self.neg_weight * mean_topk(X @ self.neg.T, self.k_neg)5 return s
That's the similarity scorer. A job scores high when it sits close to my 3 nearest saved jobs, minus half of how close it sits to my 5 nearest rejected ones.
I don't have enough swipes yet for the model to find that videographer job for me. That was the idea, though, and it gets more data every time I open the app.
Running on an old gaming PC
It all runs locally. Everything, including the embedding model, lives on an old gaming PC I turned into an Ubuntu server. I wrote about that build in a separate dev log.
Docker Compose runs the app. Tailscale gives it a real HTTPS address, which is what lets it install on my phone like a normal app. And yes, it's a bit of a flex that it's self-hosted on my homelab.
Did it work?
It actually did make job hunting more fun. I've found and saved some jobs with it, and saved jobs go into a small tracker with a stage and notes for each one.
But the bigger change is how often I open it. It's something I'd actually open now and then while waiting for something. I wouldn't do that with the normal Jobindex.
Update · Oct 2026
Most of what changed since I wrote this came from actually using it.
More jobs from Jobnet
Jobindex was missing jobs that were on Jobnet, so Jobnet is now a second source. Plenty of ads are posted on both, so the app works out when two ads are the same job and only shows one of them.
Jobnet doesn't use Jobindex's categories, which left my category filter with nothing to go on for those jobs. So the app learns it from the duplicates: when a Jobnet ad matches a Jobindex copy, it remembers which Jobindex categories go with that Jobnet occupation, and uses them for every Jobnet job with the same occupation.
From phone to PC
I swipe on my phone, but applications happen at my PC. Swipe2Job now has an MCP server, so Claude Code can list my saved jobs and read the whole ad, not just what fits on a card. It can also mark a job as applied and add a note, and notes get added to the job instead of replacing what's there.
The point is to make the jump from phone to PC easier, and to get help from Claude with the application once I'm there. It only answers inside my Tailscale network, behind its own token.
Deadline reminders
This one was on the "what's next" list, and it got bumped up. On October 15 I found two saved jobs that had closed the day before and one that closed that day, and I hadn't noticed any of them.
Now the app sends a web push at 08:00 when a saved job has 3 days left, and once more on its last day, all in one notification. A daily check spots saved jobs whose ads closed early, so the reminders only cover jobs that are still open.
Making it feel like an app
Getting the home-screen app on my iPhone right took a while. The tab bar wouldn't stay at the bottom, and the whole page could scroll and rubber-band. PWAs and phones can be their own world. A swipe also used to open the job's details by accident; that's fixed, and the card moves more freely now.
What's next
Right now it's mostly for my own job hunt. I've been thinking about making it public at some point so it can help others.
If I do, you should get to pick which job sources you connect. The Hub doesn't make much sense for someone looking for a social job, for example.
Closer to home, I might add more places to get jobs from, and rethink Saved and Applied with a search. I like the idea of a stats page too. Other than that I'll keep using it and fix the small things as they show up. Sometimes you need to use the app before you notice what needs to be added or changed.