Job boards have quietly turned into a speed contest, and one developer decided to stop losing it. Over 19 days his homemade AI job search automation setup sifted through 13,648 vacancies, sent 40 tailored applications, and never once applied for a dentist's job — even though it read plenty of them. The interesting part isn't the volume. It's the two mistakes he made along the way, and the cheap fix that saved both his money and his evenings.

What happened

If you've looked for work lately, you know the scene. You've got a browser tab stuffed with listings, ten of them skimmed diagonally, and a plan to "get back to the solid five tomorrow." Tomorrow they're gone. On a popular posting, bots fire off applications every four seconds. The author clocked one: 264 responses in the first 18 minutes. Your place in the recruiter's queue isn't decided by your résumé — it's decided by whether your application landed in the first minute of the ad's life.

So he automated it, the way programmers do. Not with the sledgehammer version, though. A "script that applies to everything" gets you a personal blacklist at a company within two days, and cover letters built from "Hello, I was very interested in your vacancy" are recognizable from a mile away. He built a sniper instead — something that reads carefully and reacts instantly, which is a combination humans can't really pull off at 2 a.m.

The one-minute loop

Every 60 seconds the system runs the same cycle. Six searches on hh.ru, the big Russian job board, plus a de-duplication pass (have I seen it? applied to it? parked it?) keyed on job title plus employer. Then a cheap pre-filter that runs before any AI gets involved: does the title contain a role token like "LLM" or "machine learning"? Does it contain a profession stop word like "doctor" or "HR"? Is the salary below his floor? Is it junior or an internship? Only the survivors get their full listing pulled and handed to a language model, which scores the fit against his résumé from 0 to 100 and explains its reasoning.

Then comes the fork. A score of 45–54 means the job gets parked and dropped into his Telegram with "release" and "skip" buttons. A score of 55-plus on a posting less than ten minutes old means it's auto-fired — letter and all.

Around that core sit the support systems: his résumé auto-bumps every four hours because hh ranks by freshness; a bot answers AI screener questionnaires in chats; a "nanny" routine reports what's happening with each application and cleans up stalled conversations. All of it is wrapped in silent error handling. If Telegram goes down, the application still goes out. If the job radar falls over, the one-minute tick isn't affected. Nineteen days in, every one of those safety nets had caught something real.

The number that rewrote the architecture

Version one was clean logic: salary gate, seniority gate, score whatever's left. Then he opened the logs. In the first weeks, 5,733 vacancies made it through to the LLM. 5,659 came back as rejections — virtually every single one of them. Dentist. HR business partner. SMM marketer. "Commercial director." He was paying for tokens so a model could read a dentist's ad carefully and politely explain that he isn't a dentist. Four to ten seconds each time.

The follow-up analysis was worse. In 71% of those cases, the title didn't contain a single token from his own profession. Not "hard to tell whether it fits" — literally "recruitment manager." You can see that from the headline alone, for free, without reading a word of the description.

That's where the token gate came from: no role marker in the title, no AI call. It's the least glamorous line of code in the entire project and probably the most valuable one.

Why no AI writes the cover letter

The one thing in the pipeline that no language model touches is the message that actually reaches a recruiter. That's deliberate. Templated letters have a smell to them, and anyone who screens candidates for a living has developed a nose for it. The filter is where the machine wins. The pitch is where a human has to show up.

What it means for you

Strip away the code and the lesson is universal: in any market where speed decides who gets seen, you either automate the boring sorting or you lose to someone who did. Here's how that plays out in ordinary life.

Scenario: You're job hunting right now

You don't need a server. You need a saved search that pings you the moment something posts, plus a five-minute triage habit instead of a two-hour scroll on Sunday night. The developer's biggest win wasn't the 40 applications — it was never again reading a listing that was never a match.

Scenario: You're coming back after a career break

Parents returning from leave, people re-entering after illness, anyone whose résumé has a gap — you're the perfect candidate for this. Your problem isn't judgment, it's throughput. Let the machine watch the board around the clock and hand you only the three listings a day that clear your bar.

Scenario: You run a small business and you're hiring

Flip the whole thing around. If bots can send 264 applications in 18 minutes, your inbox is full of noise. Write job titles that are specific enough that a keyword filter can match the right person — and expect candidates to be using AI on you too.

Scenario: You're a student or a new grad

Your first job search is mostly volume, and volume is exactly what tires people out. Set the rules once — role keywords, deal-breaker words, minimum pay — and let a simple filter do the sifting. You'll apply less and hear back more.

Scenario: You're a freelancer or creator chasing clients

Same mechanics, different board. Swap job listings for gig posts or inbound DMs, and swap "apply" for "reply in the first ten minutes with a specific first line." The person who answers fastest with a real thought about the project usually wins.

How to try it right now

Start free, and only add machinery once the manual version proves the filter works.

1. Write your profile once, in plain text. Ten lines: what you do, the tools you use, your salary floor, the job titles that are a hard no. This becomes your scoring rubric, and everything else depends on it.

2. Set up instant alerts on the job board you actually use — hh.ru, LinkedIn, Indeed, whatever your market lives on. Turn email digests off and push notifications on. Speed is the whole point.

3. Add the token gate before anything else. In your alert keywords, require at least one role marker ("machine learning", "Python", "content marketing") and exclude the professions you'll never be. This step alone threw out almost all of the junk in the developer's pipeline, and it costs nothing.

4. Score listings with a free AI tool instead of your gut. Paste the job description and your ten-line profile into a free assistant and ask for a 0–100 fit score with reasons. You can run this loop without paying a cent using the free AI tools at MyKreaTool — do twenty listings by hand and you'll quickly see which ones you'd have wasted an evening on.

5. Run two lanes, not one. Confident matches get an application immediately. Borderline ones get parked in a note or a chat with yourself, and you review them once a day with a coffee. Decide the threshold in advance so you're not renegotiating with yourself at midnight.

6. Wire up a Telegram bot when you're ready. Once the manual version works, a simple bot that pushes each shortlisted job with two buttons turns a 20-minute review into a 40-second one.

7. Write the outgoing message yourself. Keep a skeleton — one specific line about their product, one line about your relevant proof, one line about what you'd do first. No AI prose. Recruiters can tell.

8. Keep a simple tracker. A spreadsheet with company, date, score, and status. The developer's "nanny" routine exists because applications silently die in inboxes, and nobody tells you.

Upsides and what changes

Once sorting is automated, the psychology flips. You stop doom-scrolling and start reviewing a short, pre-filtered list. Your response time drops from hours to minutes, which is the only variable that matters on a crowded posting. You read fewer bad matches, so you have energy left for the good ones. And the numbers are honest: 13,648 listings scanned sounds like a machine doing your work for you, but what it really did was delete a mountain of distractions before they reached a human brain.

Limitations

This is a sorting machine, not a hiring machine. It cannot tell you whether you'd actually enjoy the job, and it gets tripped up by titles that don't match their descriptions — a "product manager" at a hospital is a very different role from a "product manager" at a SaaS company. The scoring is only as good as the profile you feed it, so a vague ten lines produces confident nonsense. Every edge case you don't anticipate (a posting in two languages, a listing with no salary, a recruiter who replies three weeks later) has to be handled or it quietly breaks something. And the letter, the interview, the judgment call about whether the company is any good — that's still entirely on you. Automating the queue doesn't automate the career.

Conclusion and one action for today

What makes this case study worth reading isn't the AI. It's that the developer measured, found that almost everything reaching his smartest, most expensive component was garbage, and fixed it upstream with a rule that cost nothing. Most people bolt AI onto a broken process and then blame the AI. Do the boring part first.

Your move today: write the ten-line profile and set one alert with a required role keyword and two hard no-words. Twenty minutes, no code, no subscription. If that alone doesn't sharpen your job search this week, you'll know exactly which part of the pipeline to fix next.