How we count
We read the open postings on 391 employers' public job boards every day and sort them into four roles by job title. Our rules find only part of the roles that belong, so every count is a floor.
Contents
What it counts
It counts open postings that employers publish on their own public job boards, and sorts each one into platform engineering, DevOps (build and release), SRE, cloud and infrastructure, or MLOps and AI infrastructure. Platform and cloud infrastructure are shown as one role.
It also reads seniority, remote work, country, the salary when the posting states one, and which skills the posting mentions.
It doesn't measure the job market. The panel leans to tech employers, mostly in the United States, that use Greenhouse, Lever, Ashby, Workable or SmartRecruiters. Big tech on its own systems, LinkedIn and Indeed are out. A count is a floor, not a total. What you can compare is the change from one month to the next, with the same rules on the same panel.
Sources
Only the public job-board interfaces that the employers' own hiring systems offer to feed their careers pages. We publish facts (title, employer, location, dates, stated salary), our classification, and the link to the original posting.
We never republish the posting's text. We read it only to classify, and it isn't stored in anything we publish.
LinkedIn isn't read. Its terms forbid scraping, so we only build links to LinkedIn's own search. Some sources were reviewed and left out because their terms forbid copying or linking their listings.
The panel
The panel is a fixed list of employers. We wrote the candidate list before reading a single posting and chose tech employers with a known careers page, not employers known to hire platform people, so nothing in the choice favours the thing we measure.
A board enters the panel if it answered and had at least one open posting. Each market has its own panel: the United States panel, one for Brazil and one for Latin America.
If fewer than 90 % of a market's boards answer on a day, that market's figures aren't published that day and the last published page stays. The postings are still refreshed, because they are facts as of the day.
Classification
The job title decides. The description only breaks ties on ambiguous titles, and the ATS department only rescues generic titles in departments named exactly Infrastructure or Production Engineering. We leave out security, corporate IT, hardware, data engineering, product, sales and support roles, and roles that build a vendor's own product platform.
Seniority comes from title markers. With no marker it stays «level not stated», which we don't treat as mid-level. Titles in Portuguese and Spanish are normalised to the English vocabulary first.
A posting that our validation samples flagged as a wrong class, or whose class the labellers disagreed on, isn't shown in the list. It is still inside the counts, which we never edit by hand.
How accurate it is
We checked the rules on 330 postings that were not used to write them. Two labellers sorted each posting without seeing our rules; they are model agents working from a written guide, not people.
| Role | Precision | 95 % interval | Recall | 95 % interval |
|---|---|---|---|---|
| SRE | 0.97 | 0.90–1.00 | 0.72 | 0.49–0.97 |
| Platform & cloud infrastructure | 0.56 | 0.44–0.69 | 0.21 | 0.13–0.38 |
| DevOps | 0.73 | 0.57–0.90 | 0.37 | 0.16–0.89 |
| MLOps | 0.83 | 0.70–0.97 | 0.33 | 0.14–0.98 |
| All four roles | 0.80 | 0.73–0.87 | 0.36 | 0.26–0.53 |
Precision: of the postings we put in a role, the share that belongs there. Recall: of the postings that belong, the share we find. Platform and cloud infrastructure are shown as one role; measured as two classes, each had a precision of about 0.47. Labellers agreed on the role with κ 0.96; the person who settled their disagreements was not blind to our rules.
Other markets, same method
- Latin America: 186 postings checked, precision 0.92, recall 0.66.
- Brazil: 180 postings checked, precision 0.97, recall 0.55.
Denominators
Every figure carries its own denominator. Roles, seniority, country, remote and employers are shares of the in-scope postings. A role's share of all postings uses every open posting on the panel's boards, of any role.
Skills are shares of in-scope postings that come with description text; one source returns none. A posting can mention several skills, and a mention isn't a requirement.
Salaries are the annual ranges postings state, per currency, never converted. A cell with fewer than 10 ranges is withheld. Change is in points of share against the previous monthly read.
Limits
- Platform and cloud infrastructure each have a precision near 0.47, so we never headline them apart.
- Recall is about a third in the United States, so counts are a floor.
- The labellers are model agents, not people, and the person who adjudicated their disagreements knew our rules.
- The panel is a convenience sample. It leans to US tech employers on five hiring systems and has no big tech.
- Salary read from posting text is a heuristic. It is base pay as posted, not total compensation.
- «Posted» dates aren't comparable across hiring systems: some report the last publication, others the creation.
Versions
- Rules
2026-10-21.6- Panel
tpr-jobs-panel-v1 · 391- Snapshot
2026-10-31T06:32:40Z- Data contract
jobs-index.v1