The sponsor register, turned into jobs
If you're an international job seeker, you can't tell which Dutch employers can sponsor your visa. The official register lists companies, but no jobs. First Job Daily matches the two every morning and shows the open jobs, free.
- Business
- First Job Daily
- My role
- Built and run it
- Since
- August 2026
- Status
- Live, free
open jobs on the site
What it is: jobs that are online today at employers on the IND sponsor register.
Why it matters
a job seeker sees which jobs are at companies that can sponsor, without checking the register by hand.
companies with open jobs, all on the sponsor register
What it is: the employers on the register that have at least one open job on the site.
Why it matters
company names on job boards rarely match the register. Fixed rules link them, so a job only shows when its employer is on the register.
Step 03, Match namesvideo runs a day, built automatically
What it is: First Job Daily runs its own YouTube channel. At 08:00 and 18:00 the system builds short job videos with an AI voice and uploads them by itself.
Why it matters
the numbers in a video come from the same data as the pages, so a video never claims more jobs than the page it links to.
Step 06, Short videosautomated tests, about
What it is: small programs that run before any change goes out and check that the site still works.
Why it matters
what a visitor sees runs on fixed rules, and the tests keep it that way. One test checks that no AI model writes the results.
GuardrailsJob and company counts from the live site, read 2026-11-16. Free to use, no sign-up.
How it runs
Seven steps, every morning
Scroll through the steps. The shape of each marker shows who is in charge; each step links to its detail below.
- AI
- Fixed rules
- A person decides
-
Read job boards
At 04:50 a job reads the job board of every registered employer, once.
Step 01 in detail -
Import the register
At 05:40 the system imports the IND register of recognised sponsors, but only when the government has updated it.
Step 02 in detail -
Match names
Company names on job boards rarely match the legal names in the register. Fixed matching rules link the two, so a job only shows when its employer is on the register.
Step 03 in detail -
Publish pages
Pages per city, per company and per city and role, plus a sponsor checker.
Step 04 in detail -
Check adverts live
Before a job gets a link, the system checks that the advert is still online.
Step 05 in detail -
Short videos
At 08:00 and 18:00 the system builds short videos with an AI voice that runs on my own computer, and uploads them to YouTube.
Step 06 in detail -
Next step: housing
Someone who finds a job also needs a place to live. Every page points housing seekers to RenterIntro.
Step 07 in detail
Step by step
What happens at each step
Guardrails
What keeps it safe
- No model in the results. What a visitor sees comes from fixed rules. A test guards this.
- Careful wording. A company can sponsor. The site never promises a visa.
- One source for every number. Pages and videos read the same counts.
- Skip when unsure. An unclear advert is left out, not guessed.
- Tests before every push. The full test suite runs before any change goes out.
Built with
Stack, in plain words
- Python with only its standard library for the data work.
- A small Node site on Vercel.
- SQLite and Redis for the data.
- One always-on computer for the scheduled jobs, with health checks.
- Local text to speech, a video renderer and YouTube upload. About 2.900 automated tests. Built with Claude Code agents.
What went wrong
What broke, and what I changed
- My own first signup got zero jobs. On 22 August 2026 I signed up myself. The system found 43 possible jobs and sent 0, because most had no official job record behind them and it won't guess. That showed me the gap: too few job sources.
- A quality filter threw out every job. On 8 November 2026 it rejected all 672 matched adverts, most of them wrongly. I rewrote each check to read adverts the way they're really written. Usable adverts went from 0 to 27.
- I changed the paid model five times in about a month. A trial, a 30-day pass, free search with a paid email, a paid sponsor list. On 31 October 2026 I made the whole site free and took out checkout and login.
For your business
What you can take from it
- AI builds, rules serve the visitor.
AI coding agents helped build it. What the visitor sees runs on fixed, tested rules.
- One source for every number.
Pages and videos read the same count, so what you publish can't drift from your data.
- A free product can feed a paid one.
Every page sends the next need, housing, on to RenterIntro.
Sources: the live site's public counts and the product's own documentation and code, read 2026-11-16.
Next caseStudyfinderAI outreachContact
Tell me what is stuck
Book 30 minutes on video, or write me a few lines about which AI tools your team uses and where it gets stuck.
consulting@mario-behrends.com