How a read works, end to end

Methodology

01What this scores

The scorer reads a public LinkedIn profile the way a startup hiring manager screens one: employers, scope, timing, and what is verifiable. It measures what the profile shows, not the person’s ability. A strong engineer with a thin profile scores low. The score changes when the profile changes.

02The data

A read fetches the public profile at the moment you run it. Each employer on it is looked up in company data for headcount, funding, and stage at the time the person joined, so companies are judged as they were then, not as they are now. GitHub is fetched and graded only when the profile links it.

03Five measured categories

Employers and scopeWhich companies, and how much they owned there

Startup fitEvidence they have worked somewhere small and early

Job tenureHow long they stay, and whether they move up while there

EducationWhether the school is a serious one for computer science

Public workCode and projects anyone can go and look at

Listed heaviest first: employers carry the most weight, public work the least. Each category is scored 0 to 100 on the rules in its footnote on the read page. A category with no evidence is skipped and its weight is split across the others. Nothing is scored as a zero for being absent.

Education is a stand-in for evidence, so it steps back as evidence arrives. Where the employment record is already strong, the school carries less and less weight, and at the top it carries none: a long record of serious work answers the question the school was being asked to answer.

04The overall judgment

An AI model scores the full profile against a fixed written rubric. The model version is pinned, so every profile is scored by the same system. The judgment also covers signals outside the five categories: location (a hub city reads in-person ready, a bare country reads remote-only, outside the US raises timezone and visa questions for the call), trajectory, and whether the career holds together.

The top of the scale is deliberately not flat. Clearing a hard hiring bar puts a profile in the exceptional band; the last few points above that are held for people who built or run the thing itself, where an engineer in the field would recognize the work independently of the logo above it. Building something from nothing counts wherever it happened, including inside a large company: founding a product that reached millions is the same evidence as founding a company.

Scoring runs at fixed settings but is not perfectly deterministic: re-running the same profile can move the score a point or two.

05The number

The five category scores and the overall judgment combine at fixed proportions, with the judgment carrying more. That blend is then lifted onto the display scale, which runs 0 to 99. A 100 is not possible: a read of a page is never the whole person, so the top is left open on purpose.

One method governs every score. When the method changes, the profiles it already ranked are re-scored on the new one rather than left on the old, so the leaderboard and a fresh read are always speaking the same language. A score is a reading taken on a date, not a permanent grade.

Rank and percentile compare the result with every profile scored so far. The percentile counts scores strictly below yours. Tied scores share the better rank.

96 to 99exceptional

87 to 95strong

80 to 86solid

61 to 79emerging

0 to 60unproven

The leaderboard is working engineers only: running a public figure through the scorer gives them a card, not a seat. Tied scores are ordered by votes. A vote never changes a score: votes are compared only after scores, so no amount of voting moves anyone past a higher number.

06The holds

Three rules bind after the judgment, on the numbers, because a model told a rule in prose will bend it for a famous name. Each fires on what the listed record shows; a section the page leaves out entirely never triggers one.

Education at 60When the page lists schools but no degree, or lists degrees and none of them is in computer science or an adjacent field, the education category is held at 60. A CS-adjacent degree at any level clears it. Only that one category is held; the other four are untouched.

Read at 74When no role on the page is a hands-on engineering role, the whole read is held at 74, beneath the solid line. Founder, executive and advisory titles are not the same as building. The card says so where the tier word would be: Held at 74, no hands-on engineering role on the profile. Linus Torvalds, Demis Hassabis and Jensen Huang all hold here, and the rule does not read Nobel citations.

Startup fitA founding role floors the startup-fit category at 85: founding is a fact, not an opinion. A history entirely inside large companies caps it at 78.

A hold is a statement about the page, never about the person. Add the role, and the hold lifts on the next read.

07The characters

Every read is also one of twelve characters, assigned from the shape of the five categories and the titles on the page. It names what kind of engineer the page describes; it carries no points and moves no rank. The card, the share image and the leaderboard all draw the same one.

Zero to OneBuilds the thing before there is a thing
In the OpenThe work is public and the record is the proof
OperatorTakes the early thing and makes it run
UpstartDoes best before the process arrives
AnchorStays long enough to see the second-order effects
MainstayFinishes what gets started
SpecialistGoes deep where most people stop
ArchitectDesigns the system other people build in
Blue ChipHas shipped inside rooms that set the bar
ResearcherComes at it from the theory first
ScholarTrained where the hard problems are set
BuilderShips the work the product actually runs on

08What a read cannot see

Private repositories, unlisted work, and anything an interview would surface. The read carries the same biases real hiring screens carry: brand names, hub cities, and clearly stated ownership count heavily. That is stated on every read.

09Zelcomp, the pay estimate

Some reads carry a pay band, which we call Zelcomp. It is not a guess about the person. It is the base salary employers certified to the US Department of Labor for the kind of role the profile describes, from about 519,000 H-1B wage filings across FY 2024 and FY 2025 (filings at exactly the prevailing wage, the legal minimum, are left out, which removes most IT-services filings), folded into cells by level, role family and metro, and by the employer itself or by companies its size when enough filings exist. Posted pay ranges are read alongside the filings: about four thousand salary bands from the public careers boards of some three hundred companies, each folded in at its midpoint. They carry the startups too small to file, where certified wages alone would run low. A cell answers only with at least twenty filings behind it, and the read names the cell it used.

Held out against employers the cells had never seen, the middle half of the band catches 49 percent of filings and the wide band 78 percent, with a median error of 14 percent (12 when the employer has its own cell). Checked cell by cell against BLS, Indeed, Levels.fyi base figures and public H-1B trackers, the employer cells sit within a few percent; SF and Seattle level cells too. Startups, and the AI labs in particular, advertise above what they certify, so a posted range can run 20 to 30 percent over this band. The total-comp line is the base band widened by what stock and bonus usually add at an employer of that size, and is deliberately wide. The career positions the band: a bare title says nothing about seniority, so the highest rank on the record and the years behind it move the band up the pool it matched, and the panel says when that happened. Engineering leadership gets a band with a caveat: at that level most of the money is equity and negotiation, so the band is a floor, sometimes priced from the staff pool, and the label says so. Founders set their own pay; their band reads as what the role pays employees. Roles outside engineering get no band.

Score your profile