Savitar · September 2026
I wanted job alerts before everyone else. So I built my own.
Watching 12,441 company career boards from a Raspberry Pi.
- median detection time
- 2.8 min
- career boards monitored
- 12,441
- postings tracked
- 508K+
A cold application sent a day after a posting goes live joins a pile. Sent in the first hour, it's near the top of it. Every job alert I tried arrived hours or a day late, so I built one that doesn't.
It watches 12,441 company career boards from a Raspberry Pi in my apartment, checks every one of them every five minutes, and has run unattended since June. A new posting is detected a median of 2.8 minutes after the company publishes it, and the alert follows within a couple of minutes.
The interesting part was never fetching job postings. It was keeping a system reliable when it depends on twelve thousand sources it doesn't control, any of which can throttle it, go down, or report a brand-new posting as months old.
Why job alerts are slow
Most startups don't run their own careers site. They rent one from an applicant tracking system, and a handful of those systems host the career boards of nearly every venture-backed company I care about. LinkedIn, Indeed and the aggregators get their copies downstream of those boards, on their own schedule.
Same posting, same company board. The only difference is who's watching it.
So the tracker isn't trying to be a better job site. It removes the aggregation delay by watching the source.
What you actually get
When something matches, it lands in your inbox. The filter is deliberately broad, so on a weekday it finds about 180 matching postings, each one emailed within a couple of minutes of being found. Here's the shape of an alert, with the postings invented:
3 new NY job posting(s):
Forward Deployed Engineer @ Northwind Labs — $170K – $210K [Series B]
Deployments · Hybrid · Full-time · posted today
New York, NY
https://careers.northwind.example/jobs/4821
Senior Backend Engineer, Data Platform @ Fernbrook Health — $180,000 – $215,000
Engineering · Remote · Full-time · posted today
Remote (United States)
https://careers.fernbrook.example/jobs/0917
Solutions Engineer @ Corvane AI — $150,000 – $185,000
Solutions · Hybrid · posted 2d ago
New York, NY, Boston, MA
https://careers.corvane.example/jobs/3380
(⚠ seen under a different posting ID 1x before at this company, most recently 2026-08-14 16:05:03)
Every part of an entry is dropped when a board doesn't supply it. The warning on the last one flags a repost, since a role posted twice is often one that didn't fill.
Want these for your own job search? I'm looking for a few people to try it first.
How fast it is
The tracker detects a new posting a median of 2.8 minutes after the company publishes it. That's measured on 64,533 postings published between August 29, when the current five-minute schedule began, and September 26, 2026, by comparing the platform's own publish timestamp with the moment the tracker first recorded the job, on the two platforms that report a reliable publish time. Postings whose publish time changed after the tracker saw them, and re-opened roles that keep their original date, are left out. Platforms are anonymized.
| Postings | p50 | p90 | Within 10 min | Within 30 min | |
|---|---|---|---|---|---|
| Platform A | 13,405 | 3.2 min | 5.2 min | 98% | 99% |
| Platform B | 51,128 | 2.8 min | 4.9 min | 95% | 96% |
| Both | 64,533 | 2.8 min | 5.0 min | 96% | 97% |
Platform B's tail is the throttling bug described below: until September 22, a board it throttled was switched off for up to a week. In the four days after the fix, 99% of postings on both platforms were detected within 10 minutes, and the slowest 1% within 7.
To see what those minutes buy, I checked LinkedIn directly. On one afternoon in September I took 17 postings the tracker had seen two to eight hours earlier and looked each one up. Eleven were listed, a median of about an hour and a half after the tracker saw them, and two of them four to five hours after; one was a tie, a company that posted to LinkedIn at the same moment. By the time I looked, 4 to 44 people had already applied or clicked apply on each listing, and one had drawn 39 within about two hours of appearing. I couldn't find the other six, mostly AI startups, on LinkedIn at all, even three to six hours later. LinkedIn shows only the hour a job was listed, so each gap is good to about half an hour, and 17 postings is a small sample.
The next afternoon I ran the same check on Indeed and Google Jobs with 18 fresh postings. Indeed records the minute it adds each job: it had 7 of the 18, every one after the tracker, a median of about two hours later. Google Jobs had just one, a role the company had apparently posted to LinkedIn two days before it reached its own careers page, the only posting in either check the tracker didn't see first.
For additional context, a published test measured LinkedIn's alert emails at 31 hours after posting on average. The paid alert tool that ran the test measured its own at 8 minutes.¹
The difference between a few minutes and 8 is probably not meaningful. The difference between minutes and a day is the product.
How it watches 12,441 companies
Discovery runs weekly. Everything after it runs every five minutes.
None of those applicant tracking systems publishes a list of its customers, so the tracker builds and maintains its own: 12,441 live boards, monitored continuously.
Every five minutes the tracker checks every known board. That's 12,441 boards × 288 cycles, about 3.58 million board checks a day from one Raspberry Pi.
Before the speed work, a cycle over 11,345 boards took 424 seconds. Two changes brought the core poll down to 29 seconds; the rest of today's 135 seconds is spent on purpose.
| Stage | Cycle time |
|---|---|
| Before: every board fetched in full | 424 s |
| Boards answer "unchanged" in one round trip | 176 s |
| HTTP client with a C parser | 29 s, the core poll |
| Deliberate rate limit on one platform | ~110 s |
| Today: full production cycle, fifth platform included | ~135 s |
Most of the speedup came from the boards themselves: almost none change in any five-minute window, so more than 99% of the boards that can answer "unchanged" do, with nothing to parse. The rest was a CPU problem disguised as a network one. On the Pi, the old HTTP client spent about 25 ms of CPU per request and pegged a core; one with a C parser does the same request in about 2.
What breaks
Fetching one job board takes a few lines of code. Everything around it is the actual system:
- Which companies exist? No platform publishes a list of its customers.
- Is this board still alive, or did the company move to another system?
- Is this posting new, or a re-opened role that kept its old date?
- Is it a repost of a role that didn't fill?
- Is it one posting, or the same posting listed once per office?
- Did that request fail because the board is gone, or because I'm being throttled?
Each of those has been answered wrong at some point, and every wrong answer either floods the inbox or silently drops a job. Three of those failures, from the logs:
Throttling that looked like deleted boards
- What happened. One platform's CDN rate-limits by IP and answers with a plain 403, no Retry-After. The tracker read any 4xx as "this board no longer exists," and it stamped failed boards with the newest timestamp, which sorted them to the tail of the next cycle, exactly where the throttle bit again. The log showed streaks of exactly five cycles with a near-constant failure count (34, 34, 26, 35, 35), then zero.
- Why it hurt. Five strikes switched a board off until the weekly discovery pass revived it. Two passes in a row reactivated 335 and 285 boards: about 300 live boards a week sat blind for up to seven days, and about ten matching postings a week reached me days late.
- Fix. Failures now split into gone and transient, and only a definitive "does not exist" counts toward switching a board off. A transient failure writes nothing, so the board is retried at the front of the next cycle instead of the back. A rate limiter keeps that platform under its budget.
- Result. In the first 370 cycles after the fix, not one request was throttled, and only four cycles saw any transient failure.
The 29-hour freeze
- What happened. The Pi's WiFi dropped in the middle of sending an alert email. The OS never errored the socket, so the mail client waited forever, while holding the lock that keeps runs from overlapping.
- Why it hurt. For 29 hours every scheduled run found the lock held, assumed a run was in progress, and exited quietly: no alerts, no errors, just one line per skipped run in a log nobody was reading.
- Fix. A timeout on every mail connection. A watchdog that aborts a stuck run through the normal failure path: one "tracker is BROKEN" email, lock released. And a cable instead of WiFi, whose flaky DNS was behind nearly every "suspected outage" in the log.
- Result. A hung run now ends itself and says so, instead of silently holding the whole system for a day.
Re-opened roles that looked months old
- What happened. A freshness rule stops a newly discovered board's backlog from being emailed: anything published more than 14 days ago is skipped. One platform keeps the original publish date when a closed role is re-opened, so a re-opened posting looked months old.
- Why it hurt. About 4% of matches were silently dropped, one of them a software engineering role at Anthropic in New York.
- Fix. The age rule now applies only to a board that hasn't had a successful check in the past day, including one being checked for the first time. On a board checked five minutes ago, a posting the tracker hasn't seen is new by construction, whatever date it carries.
- Result. Re-opened roles arrive like any other posting, and a new board's backlog still stays out of the inbox.
What makes the alerts useful
The filter favors recall over precision, on purpose: a false positive costs a few seconds of reading, while a false negative can mean missing the job entirely. It understands New York and its boroughs, explicitly US-remote roles and a broad set of engineering titles. Adding the US-remote rule alone roughly doubled the matches: about 450 engineering postings a week that a New York-only filter had been dropping.
Pay and funding stage come from the posting wherever it states them, and New York's pay transparency law puts a range in nearly every New York posting. Both are metadata for skimming, not authoritative company data. Reposts are flagged.
The dataset
Every posting ever seen is kept, 508,815 so far, and that's what makes reposts recognizable. The alert system accidentally became a historical record of hiring. It knows when each posting went up on any board it was already watching, so it can already answer questions like these:
- Which companies are hiring right now, and how fast?
- Which ones just started hiring?
- Which roles keep getting reposted?
I haven't built anything on top of it yet.
Early access
The tracker is fast enough now. The interesting question is whether anyone else wants the same unfair head start.
The expensive part already runs around the clock, and it already sees postings far beyond New York. You don't need another tracker, just another filter over the same stream: your titles, your city, your inbox. I'm looking for a small group of job seekers to try the first version. Email me the roles and the city you're after.
Building something where minutes matter? I'd like to hear about it too: LinkedIn or GitHub.
Technical detailsThe latency breakdown, the discovery outage, three more failures, and what I learned.
Latency by platform
Both numbers are essentially the schedule: each board is checked every five minutes, so a new posting waits two and a half minutes on average, and Platform A's own cache adds up to about a minute on top. The median stayed between 2.8 and 2.9 minutes in every full week measured. The slowest 1% took up to 41 minutes on Platform A, and on Platform B about 15 hours before the throttling fix and 6 minutes after it. Platforms that record only when a posting was created, or no time at all, are left out.
The company list
A weekly discovery pass over a public, web-scale dataset produces candidate boards, and each one is checked against the live board before it's stored. The list holds 12,441 live boards, and the weekly pass now adds only one to three a week. An outside check says it's close to complete for the companies I care about: of the boards on those systems linked from three months of Hacker News "Who is hiring" threads, 87% were already on it, and the few live ones it lacked had no matching roles.
The list didn't start saturated. One deeper discovery pass in August found 1,466 live boards the weekly pass had never seen, Mistral AI, Whoop, Wayve and Riot Games among them. At least one live board never appears in the public dataset at all and was added by hand.
That same month the public dataset's query service died and stayed down for weeks. Nothing alerted: polling kept delivering from the known boards while the company list silently froze at 5,204. The service turned out to be a thin layer over static files that were still being published, so the tracker now reads those directly and depends on nothing but object storage. A discovery source that's been dead for two weeks now puts a warning in every alert email.
More failures
The run the OOM killer took
- What happened. On the first full baseline, the run died with no traceback and no failure email. Fetches were outrunning the database writes, and completed but unprocessed responses piled up until the OOM killer took the process. One board fetched with descriptions can be tens of megabytes; the biggest I've measured are 29 MB and 41 MB.
- Why it hurt. A killed process can't report its own death. The only symptom was a cycle that never finished.
- Fix. Capping requests in flight isn't enough, because the memory lives after the fetch. Each worker now fully processes a board before fetching its next one, which caps parsed responses at the worker count. That ordering is the backpressure.
- Result. Peak memory is bounded by the worker count, whatever the boards weigh, and a cycle cut short resumes with the boards it never reached.
One row per location, not per posting
- What happened. The fifth platform lists a multi-location posting once per location, under the same id. One veterinary chain's board had 569 rows for 273 jobs.
- Why it hurt. The tracker de-duplicates by id, so the first version only ever judged the first location listed. 56 of the 457 matching postings in stock would have been missed, every one with its New York or US-remote location listed second or later.
- Fix. Rows are merged into one posting that carries every location.
- Result. Caught on the platform's first live cycle, before a single match was lost.
The 15-minute lockout
- What happened. One platform's per-board API has no "unchanged" answer, and a short burst of requests earned a 15-minute IP-wide lockout with no Retry-After.
- Why it hurt. Polling its boards one by one wasn't viable at any useful speed.
- Fix. The platform turned out to run a public search across all of its customers' postings, so the tracker walks its New York results instead: about a dozen requests a cycle.
- Result. Every New York posting from that platform's customers in about two seconds a cycle, including boards no discovery pass ever found.
Plumbing
Around the poll sit incremental commits, oldest-checked boards first so an interrupted cycle resumes where it starved, and a durable queue of matches that's only cleared once the email is confirmed sent. The alert email's footer carries diagnostics, because it's the one channel that still works when the Pi won't take an SSH connection.
What I learned
- Measure before adding coverage. When I added a third platform, its 1,805 boards turned out to add about 40 matching postings a week against about 850 from the first two, while the latency work that same week cut the cycle 14x. Every platform since has been measured with a throwaway poll first.
- A transient failure must never look like a deletion.
- Bound the work you've parsed, not just the requests in flight.
- On a small machine, the CPU can be the bottleneck of a network job.
- Check whether a platform publishes its own feed before polling its boards one by one.
- Check whether a list is one row per posting or one row per location before trusting its counts.
- Silence is the worst failure mode. Anything that can fail quietly needs something watching it. Here that's a watchdog on every run, a warning when discovery goes stale and a warning when a feed stops answering.
Sources
- ¹ The same test measured Indeed's alerts at 18 hours, Glassdoor's at 28 and Google Jobs' at 42, all averages. Developer job-data feeds mostly state refresh cycles of 1 to 24 hours.
- Job alert speed test, 50 postings: the LinkedIn, Indeed, Glassdoor and Google Jobs alert figures and the paid-tool figures
- LinkedIn listing times and apply counts: read from LinkedIn job pages, September 2026
- Indeed and Google Jobs listing times: read from their search results, September 2026
- Feed freshness: the vendors' own API pages, September 2026