Job boards hand you a search box and a few thousand listings, and leave the rest to you. You do the checking, the reading, the judging and the writing — every time. And they only ever show you what exists right now. Nobody tells you what changed.
A role you have been tracking for three weeks quietly adds a requirement you do not have. A posting gets rewritten from “Data Scientist” to “Principal Data Scientist” and the experience bar moves from five years to twelve. The job you applied to last Tuesday is pulled on Thursday and nobody says a word.
You would only notice any of this by re-reading every posting you care about, every day, and remembering exactly what each one used to say. Nobody does that. That gap is what this system fills.
Today a job search means doing four tedious things by hand, over and over: checking every company’s board, reading each posting to see if it is even relevant, judging whether you are actually a fit, and working out what to say if you apply. Jobs NightWatch does all four, on a timer, and hands you the answer.
It filters against your actual profile. Of 2752 postings across 10 companies, only 295 match your background — your target titles, your real skills, your seniority. You never see the other 2457.
It rates what is worth applying to. Every change that matters gets a fit score out of 100, with the reasoning behind it. Not a keyword count — an actual assessment of whether your experience lines up with what the role is asking for.
It drafts what you would lead with. For roles worth your time, it writes the specific bullets to open with, citing real projects and real numbers from your background — so the blank page is already filled in before you start applying.
And because it knows what each posting said last time, it can tell you things a job board structurally cannot: that a role you were interested in just raised its experience bar past where you sit, or that the team quietly stopped hiring.
Jobs NightWatch watches 10 companies’ career pages on a timer. It remembers every posting it has seen. When something changes, it reads the posting properly, works out whether the change matters to you, and writes up what you would say if you applied.
Cloud Scheduler wakes the system every few hours. This is the difference between an agent and a website you have to remember to visit. You can also trigger a run yourself with the Run now button.
Every posting is normalised and fingerprinted with a SHA-256 hash of its content. Today’s fingerprint is compared against the one stored last time. That comparison is plain deterministic code — no AI. Deciding whether two records differ is a comparison, not a judgement; using a language model there would be slower, more expensive, and less accurate than a hash.
Of 2752 postings being watched, 295 match your profile — a deterministic check against your target titles and core skills. This is a cost guard, not the fit decision. A warehouse role does not need a language model to rule it out.
Only changes that survive the filter reach Gemini 3.7 Flash. The agent decides which tools to call: it fetches your profile, retrieves what the posting said last time, runs a deterministic eligibility check, and records its verdict. The model does all the judging; the tools only fetch facts and save results.
It is not a job board. The company view shows which roles match your profile, but the model is only ever spent on roles that changed. Scoring all 295 matching roles would cost hundreds of model calls to tell you nothing you did not already know.
It is single-tenant today. One profile, one set of tracked companies. The data model was built to allow per-user profiles later, but that is not built yet.