An agent that finds the jobs. I make the decisions.
It watches 47 companies' job boards, scores every role against a CV with an explainable breakdown, and prepares a tailored CV and cover letter on request. It never applies or messages anyone by itself.
Try it: the real flow, on your phone
I use it through Telegram. Below is that exact experience, with a fictional candidate run through the real pipeline against real, public job boards. Tap the buttons, or watch the whole flow replayed.
Uploading your own CV works in the live version of this app (it runs the same code and calls the model per visit, so it is not open to the public). This page replays a recorded run of the real pipeline so you can see the whole flow.
Use your own CV
Paste text instead
Your CV is processed in memory for this session only, sent to Anthropic's Claude API for analysis, never stored, and erased after 60 minutes.
Candidate profile (the only thing that changes per person)
Sample CV the agent works from (fictional)
The live version (upload your own CV) runs the same code. It is not open to the public because every visit calls the model, and CVs are personal data.
How it works
A pipeline of small, single-purpose agents with one deliberate human gate. Rules do the cheap work; the model is used only where judgement is needed.
- Job boards9 hiring systems, public APIs only
- ScoutFilters, dedupes and tracks what is still open
- AssessorLLM 5-part score with hard filters
- ReportEvery open job, new ones marked
- Human gateI decide and tap a button
- TailorOne-page CV and a cover letter
Design decisions
The interesting part of an agent is rarely the model call. It is what the system is allowed to do, and what happens when it is wrong.
Human in the loop by design
The agent drafts; I decide. Buttons generate a CV or letter on demand, and applying, messaging recruiters and sending anything stay manual. Autonomy is a product decision, not a default.
Grounded generation, not free writing
The CV is built from a verified facts file with a "do not claim" list. Names, titles, dates and education are fixed in code, so the model only writes bullets. A real catch: it once turned "probably Dynamics 365" into a stated fact, so the rule now forbids stating inferences.
A score you can argue with
Five weighted components with strengths and gaps that must cite the CV. Director-level roles are a hard filter, so other components cannot rescue them.
Failures never hide a job
A broken job board does not stop the run, and its jobs stay in the report. If a score fails, the job shows as "not assessed" instead of vanishing. Reports always list every open match, so the output is a state, not a diff.
Rules first, model second
Title, location and age filters run for free. The model scores only survivors, once per job. No new jobs means no report and no model calls.
LinkedIn scraping and auto-messaging
Scraping breaks the platform's terms and risks the account; automated outreach damages the relationship it is meant to start. The agent uses companies' public job APIs instead.
Limits and what is next
Scores are not calibrated yet
They come from model judgement against a written rubric. Next: compare them with human judgement on a labelled set and report agreement.
Coverage is partial
It covers companies whose job boards expose a public API. Sites that render everything in the browser are not covered yet.
Some posting dates are approximate
A few boards only expose a last-updated time, so a refreshed old job can look newer than it is.