What JobRadarly Can and Cannot Do, and the Limits You Will Hit

The limit is shared, and that changes what it means

This is the single fact most visitors would want to know first, and it is the one no other page on this site says plainly. The request ceiling is not per visitor. It belongs to the server process, and every caller spends from the same budget.

The reason is not stinginess. The ceiling exists to protect the three boards this site reads, and it deliberately does not key on an IP address or any request header. Behind a reverse proxy those values are supplied by the caller unless a trusted proxy overwrote them, so keying the limit on them would let anyone escape simply by rotating the value. A limit that the limited party can rewrite is not a limit.

The ceilings, in one table

The fourth row is a different kind of bound and it is the one people miss. A rate limit caps how many operations start in a window; it says nothing about how many are open at the same time. Six calls admitted in the first millisecond each hold a provider connection at once, so a rate limit alone would not bound the simultaneous work. A separate gate does, and when it is full the request is refused before any source read or provider call begins.

Three companies, and no way to add a fourth

JobRadarly reads Figma, GitLab and Robinhood, from their own boards on Greenhouse. That is the entire reachable set. There is no parameter for a company, a board token or a URL, and the request path is assembled from a constant in the source rather than from anything a visitor sends.

This is a deliberate refusal to be useful in a particular way. Supporting any company would mean accepting a board name from the caller, and a route that fetches whatever host a stranger names — from inside the server, under the site's own address — is an open proxy with a job-board interface. Three named boards is the price of not being that.

Those counts move, because they are postings rather than constants. At the last check the three boards carried roughly 154, 223 and 154 postings respectively; a number quoted in a document would be wrong within a week, which is why no test asserts one.

What each answer actually means

The live job API separates facts that are easy to collapse into one message, and the separation is the useful part. A posting the source has stopped advertising and an address that was never a posting at all are not the same problem and do not have the same remedy. Reporting the first as the second tells a reader their link is broken when it is not.

On a single posting

On a whole board

The asymmetry with the single-posting case is intentional. A 404 on one posting is an answer about that posting. A 404 on a whole board is a failure, because an unreachable board tells you nothing about whether a company has open roles, and reading it as zero postings would be inventing an empty list.

The same rule covers a subtler case: a board that answers 200 with a body that is not a board at all — a maintenance notice, a bare list, a null where the postings belong. That is a failed read, not a board advertising nothing, and it leaves the previous postings exactly where they were. Before that distinction existed, a maintenance page could make the page show fewer jobs than a moment earlier, dressed up as freshly read.

Every way the AI review can fail, and who is responsible

A paid feature that fails owes the person who paid an explanation, so the AI route does not have one generic error. It has twelve distinct reasons, each mapped to its own status code and its own sentence. The table below is the whole set.

Two things in that table are worth calling out because they are where most products quietly lie. First, throttled returns the points reserved for the request in full, because no analysis was attempted. Second, a failure after the provider was reached also returns the reservation in full, because nothing was delivered. The charge is what came back, not what was asked for.

And a review that concludes not a fit does cost a point. That is not a loophole in the refund rule; the work was done, and it is evidenced, which is exactly what was bought. What costs nothing is an answer that never arrived.

Five things this site will never tell you

Boundaries are more useful when they are named than when they are merely absent. These five are the ones a reader is most likely to assume exist.

• A salary number the posting did not state. The only salary evidence a live record may use is the pay range the board itself supplies. Nothing is estimated, averaged or filled in from a similar role.
• A list of every company that is hiring. Three boards are read, named in a constant. The reachable set is a line of code somebody can read, not something a request can widen.
• A sponsorship verdict. The site shows which roles survived your filters and why each one survived. Whether an employer has a sponsorship record is public and you check it yourself.
• A match score out of a hundred. The AI returns one of four labels with excerpts, never a number. An invented number is not a measurement.
• A guarantee that a posting is still open. A posting leaving a board's list is evidence that it left the list, and nothing else. The page never says a role was filled, because that cannot be known.

What is kept where

There is no database behind the job list and no queue and no timer. The cache is a map inside one running server process. That is the whole persistence story, and it has two consequences worth stating rather than hiding: a restart forgets when each posting was first observed, and a second server process would have its own separate view.

Your profile, your saved jobs and your application tracker are not on the server at all — they stay in your own browser. What does live on the server is your points balance and any analysis you paid for, so that re-reading it is free and survives a cleared browser. That split is why buying points requires a signed-in account while browsing does not.

How a bad day at one board is handled

Each board gets fifteen seconds before the request is abandoned. Within the ten-minute cache window no request goes out at all; past it, the next read does. A click that arrives while a refresh is already running joins that refresh instead of starting a second one, so an impatient double-click cannot become a burst.

Frequently asked questions

How many requests can I make?

Ten board refreshes per minute and sixty single-posting reads per minute. Both ceilings belong to the whole server process, not to you, so another visitor's traffic counts against the same budget. Normal browsing never comes close: a page load is one refresh and a few detail reads, and a cached read does not count at all.

Which companies can JobRadarly show jobs from?

Three, and only three: Figma, GitLab and Robinhood, read from their own boards on Greenhouse. There is no parameter for adding a company. Support for any company would mean accepting a board name from the caller, and a route that fetches whatever host a stranger names is an open proxy.

Why is the request limit shared instead of per user?

Because the limit exists to protect the three boards we read, not to meter individual visitors, and the server cannot trust a caller-supplied IP header. A forwarded header is client-controlled unless a trusted proxy overwrote it, so keying the limit on it would let anyone escape by rotating the value.

What does 404 mean on a job page?

That the source answered and the posting is not in its list any more. It is an answer, not a failure. The response also carries a reason: not-ours for an address that was never a posting on the whitelist, not-listed for one the source has stopped advertising, and unreadable for a reply that was not a posting at all. None of these means the role was filled.

What does 502 mean on a job page?

That the source did not answer, so we do not know whether the posting is still live. This is deliberately a different status from 404, because the two are different facts and only one of them is about the posting.

How long is a job list reused before it is read again?

Ten minutes. Within that window the page answers from memory and does not touch the boards at all. Past it, the next read goes out. Cached reads are cheap and are never refused by the refresh limit, which is the point of the split: the limit bounds outbound work, not reading.

What happens when one of the three boards is down?

The other two still answer, and their postings still show. The failing board's previous postings stay on the page carrying the time they were actually fetched, marked as older than this attempt. Nothing re-dates an old list as a fresh one.

Can JobRadarly tell me whether a company will sponsor me?

No, and no tool can. It reads the companies' own boards and shows which roles survived your filters, with a reason for each. Sponsorship history lives in public filing data and you check it yourself. Any tool that implies otherwise is selling certainty it does not have.

How many jobs can I analyse with AI at once?

Ten per request, and you pay only for the analyses that come back. A job with no description, an analysis the provider never answered, and an answer that failed validation all cost nothing. An analysis concluding not a fit still costs one point, because the work was done and it is evidenced.

Does the AI return a match score or a percentage?

No. There is no numeric score, percentage or confidence field anywhere in the review, deliberately. A number invented by a language model is not a measurement. You get one of four labels, plus strengths, gaps and unknowns, each backed by an excerpt that must appear in the posting text.

What is the daily short list based on?

Your own constraints, applied to what the boards currently advertise. The result is sometimes three jobs and sometimes zero, and zero is a real answer rather than an error. A search page returning four hundred results is doing its job as a search engine and failing you as a decision tool.

Do saved jobs and my profile live on your server?

No. Your profile, saved jobs and application tracker stay in your own browser. What lives on the server is your points balance and any paid analysis you bought, so that re-reading it is free and survives a cleared browser.