250 CVs for one Gulf posting. Not one of them leaves the machine.
Paste the posting, add the CVs you were sent, and get them back in order with the reason next to every name. Extraction, the CV check, the matching and the score all run inside this browser tab.
No CV text, no candidate name and no filename appears in any request Meritline makes — and the network policy that enforces it is published in full, so you can check that sentence before you believe it.
One card per candidate. This is an actual one.
Rendered by the same component the results page uses, from the same sample CV the demo ranks, through the same matcher. Not a picture of the product — the product.
rakesh_menon_sample.txt
Matches 15 of 15 requirements you listed.
From the CV
“- Led incident investigation on 14 recordable events and drove the corrective actions to closure.”
We publish the benchmark we failed.
Before measuring anything we set a floor: a rank correlation of 0.90 or we would not call it good. Here is what it actually measures, generated by our build from the benchmark baseline — nobody types these digits by hand.
On a 252-CV benchmark across 7 job descriptions, Recall@20 is 0.807 (95% CI 0.714–0.850), Spearman rho is 0.708 (0.650–0.747) and Precision@5 is 0.257 (0.143–0.486). That rho fails the 0.900 floor we set ourselves, and we have not moved the floor. The reference ranking is constructed by the corpus generator rather than written by people, so it measures agreement with a specification, not with hiring judgement. That is why results are grouped into bands rather than presented as an exact order.
We did not clear the floor and we have not moved the floor. It is also why a card shows Strong, Possible or Weak rather than a leaderboard: at a Precision@5 of 0.257, “#3 above #4” is not a finding, and a tool that presents it as one is lying with a sort order.
250 applicants are usually not 250 people.
The same man, three times
One CV arrives from the job board, from a sub-agent, and from the candidate on WhatsApp. Meritline reads every file on your device before anything is scored, matches exact copies by hash and reworded copies by shared phrasing, and shows you the groups.
Nothing is deleted and nothing is auto-removed. Which copy is the real one is a decision about a person, and it stays yours.
What we cannot do yet
A batch is 50 CVs, because the work happens on your machine rather than on a server we rent. A 250-CV posting is five batches today, and reuniting them into one ranking is the next piece of work.
We would rather tell you that here than have you find it on the upload screen.
When the client asks why you didn’t send him.
Every candidate you added stays on the list. Meritline puts them in an order and shows its reasons; there is no auto-reject in the product and no code path where the software turns anybody down.
Against each name you get the requirements your posting named, split into the ones the CV mentions and the ones it does not, and the line from the candidate’s own CV where each one appears. Add your own reason in a sentence, and it travels into the CSV and the PDF you send on. Run the same posting and the same files again and you get the same order, so months later you can show how a shortlist was produced instead of remembering it.
It reads the requirements a Gulf posting is actually screened on.
Safety and site
NEBOSH IGC, IOSH Managing Safely, OSHAD, permit to work, confined space entry, work at height, scaffold inspection, method statements, risk assessment, incident investigation, HSE audits, toolbox talks. Primavera P6, BOQ and quantity take-off, AutoCAD, shuttering and formwork, setting out, MEP coordination, snagging.
Facilities, MEP and GCC finance
HVAC and chillers, LV switchgear, VFDs, distribution boards, fire alarm systems, planned preventive maintenance, CMMS. VAT returns, IFRS, month-end close, SAP FICO, bank reconciliation, ACCA, CMA.
What it does not read yet
“ARAMCO-approved”, GCC country-years, trade-test results, and visa or iqama status. Those are four of the most common lines on a Gulf screening sheet and Meritline cannot see them today. They are next. Until they ship, this page says so.
For the person who has to sign this off.
Your client’s procurement pack asks three questions. Here are the three answers, each one checkable.
- Where candidate data goes
- Candidate CVs never leave your computer. Nothing is uploaded. Signing in sends your Google identity. Recording a decision sends a score id and the action — never CV text, names, or filenames. Filenames count as personal data: CVs arrive named Firstname_Lastname_CV.pdf, and none is ever put in a URL or a request.
- What is running
- all-MiniLM-L6-v2, INT8 ONNX, served from this origin rather than a public CDN, and SHA-256 checked before the scoring session is created. A tampered model is refused, not silently used. There is no language model, no sampling and no temperature anywhere in the score path.
- What it is allowed to talk to
- The content security policy, as served, is published in full. No analytics, no font CDN, no error reporter, no session replay, no third party except Google when you sign in.
Run it inside your own network.
Meritline builds to static files. There is no scoring server to run, no database to secure, and nothing that phones home to count seats — which is why it can sit on a machine in your office, or on your client’s. The first run downloads the model once and caches it; after that the ranking itself needs no network at all.
Licensing for agencies is annual, per office. Tell us how many postings you run in a month and which client’s data rules you have to satisfy, and we will tell you the number.
It sorts one batch against one posting and shows why. It does not decide who to hire, it does not run your pipeline — it hands off to whatever you already use — and it cannot find candidates who have not sent you anything.
Exactly what does and doesn't leave the device
Candidate CVs never leave your computer. Nothing is uploaded. Signing in sends your Google identity. Recording a decision sends a score id and the action — never CV text, names, or filenames.