SWE Genie

Methodology

How the taxonomy was built, how scoring works, and what it doesn't capture. Every claim on this page is traceable to a sourced research brief — this page exists so you don't have to take the result on faith.

1. Where the 18 archetypes came from

Companies verified
744
Postings classified
67,956
Archetypes with 200+ postings
13 of 18
Top-coverage role
6,208

postings · 542 companies

Each archetype started from a research brief built from real job postings across company sizes, practitioner blog posts and talks, and comp data (Levels.fyi, Glassdoor). Briefs cite every trait claim back to a specific source — no claim is asserted from stereotype alone.

A 2026-07-11 sourcing pass then verified those companies directly against their public job-board APIs and classified those postings into the 18 archetypes — not scraped or estimated, every posting traceable to a live URL at the time it was harvested. The other 5 (Embedded/IoT Engineer, Solutions Architect — Consulting-side, Developer Relations, Technical Product Manager, Solutions Architect — Vendor-side) came in between 57 and 148 postings — still several times the original sourcing bar, but a genuinely smaller public posting pool for these titles than for the other 13.

2. The 22 trait dimensions

22 dimensions10 skill · 12 preference

Every archetype is scored against the same 22 dimensions — 10 skill/experience dimensions and 12 preference/temperament dimensions. Each dimension was derived from patterns across multiple research briefs, not written top-down before the research. The taxonomy separates skill axes precisely enough to avoid conflation — ML aptitude is scored apart from general domain fit, and private mentoring is scored apart from public visibility, so overlapping traits don't get flattened into one blurry axis.

Ambiguity Tolerancepreference
Interrupt Tolerancepreference
On-Call / Incident Appetitepreference
Debugging / Diagnostic Depthskill
Systems Design at Scaleskill
Technical Breadth vs. Depthskill
Outcome Accountabilitypreference
Stakeholder & Client-Facing Comfortpreference
Teaching / Explaining Enjoymentpreference
Public Visibility Comfortpreference
Account Portfolio Breadthpreference
Relationship Continuitypreference
Variable Compensation Appetitepreference
Coding Intensityskill
Travel / Physical Embed Willingnesspreference
Adversarial / Threat-Model Thinkingskill
Physical Constraint Engineeringskill
ML Engineering Fluencyskill
Mobile Platform Fluencyskill
Data Infrastructure Fluencyskill
Cloud Infrastructure Fluencyskill
People-Outcome Management Orientationpreference

3. How scoring works

1Target + weight per dimension

Every archetype has a target value (1-5) and a weight (0-1) for each dimension, drawn from its research brief. Your answers produce a 1-5 value per dimension.

2Score your answers against the targets

Fit is a weighted average of how close your answers land to each archetype's targets, weighted by how much that dimension actually matters for that role — not a raw distance across all 22 dimensions equally. Closeness is scored with a non-linear penalty, so a large gap on a dimension counts far more heavily than a small one.

3Defining-trait floor

A role's overall score is capped by how well you match its single defining trait — its highest-weighted dimension — so no role can reach the top of your list on incidental agreement while you mismatch on the thing that actually makes it that role.

4Correlated-pair tie-breaker

Some traits reliably travel together across the 18 roles — depth at cross-layer debugging with appetite for on-call incidents, comfort teaching with comfort being publicly visible. Those pairs are measured directly from the taxonomy (a committed script, not a hidden model), and when your answers on a correlated pair line up the way a role's own profile does, that coherent combination nudges its score up; when they pull in opposite directions it nudges down. It's a deliberate tie-breaker between close matches — it can reorder neighbors but never override the defining-trait cap above it.

5Why matched / growth areas

The “why you matched” explanation is not written after the fact — it is the 3 dimensions that contributed the most to your specific fit score for that archetype. Growth areas are the dimensions pulling your score down the most, above a floor that filters out noise. There is no black-box model here: every number on a result page is reproducible from your answers and the public taxonomy data.

4. What this doesn't capture

Non-linear gap penalty
Fit uses a non-linear (squared) gap penalty: a 2-point gap on a dimension costs four times as much as a 1-point gap, not twice. Small mismatches are treated as near-noise while large mismatches count sharply against a role. One side effect: because small gaps are forgiven, the headline percentages read a little generously, so a match is best understood relative to your other ranked roles rather than as an absolute grade.
Dimension correlation layer
Some traits reliably travel together, like appetite for on-call incidents and depth at cross-layer debugging, or comfort teaching and comfort being publicly visible. By measuring which dimensions rise and fall together across all 18 archetypes, the scorer recognizes those clusters: when your answers on a pair of correlated traits line up the way a role's own profile does, that coherent combination counts as extra signal for it, and when they pull in opposite directions it counts against. It stays fully transparent — the correlations are derived from the public taxonomy by a committed script, not a hidden model, and your result page names the exact pairs that helped or hurt. It's a nudge by design: it can reorder roles that were already neck-and-neck but never override the defining-trait check above it.
Defining-trait floor
An archetype's score is floored at how well you match its one truly defining trait (its highest-weighted dimension) — so an archetype whose incidental dimensions happen to match your answers can't float to the top on broad agreement while you mismatch on the trait that actually defines it. This is enforced structurally, not left as a caveat.
Expert-authored weights
Weights are set by expert-authored research briefs, not crowdsourced averages. A short survey on your results page and a “Rate this role” control on each archetype page let people who actually do (or hire for) a role rate what it really takes — those signals accumulate per role and only factor in once there's enough volume to aggregate safely, and any resulting diff between expert and crowd input is published, not hidden. See the roadmap below.
Two dimensions we deliberately excluded
Two candidate dimensions were evaluated against the job-posting corpus and deliberately not added. AI-coding-assistant fluency (Copilot/Cursor/Claude Code as a day-to-day workflow expectation) is real and growing, but it shows up at a broadly similar rate across nearly every engineering archetype — a trait that doesn't vary much between roles gives a ranking model nothing to discriminate on, so it's addressed in role write-ups instead of as a scored axis. Government/federal security-clearance eligibility is a strong, concentrated signal in a few archetypes (Forward-Deployed Engineer especially), but it's closer to a binary eligibility fact than a point on a 1-5 preference or skill spectrum, so it doesn't fit this model's shape either — it's called out directly in the relevant role pages instead.

5. Compensation data

Company tiers
5

AI labs · FAANG/Mag7 · high-growth public · growth-stage private · early-stage

Career levels
4

L3 · L4 · L5 · Staff

Percentile bands
5

P10 / P25 / P50 / P75 / P90

Confidence tiers
3

high · medium · low

The comp-by-tier charts on each archetype page (and the Staff cross-role ranking) are a separate, hand-curated dataset from the salary ranges in the research briefs. They cover all 18 archetypes and break total compensation into base, bonus, and annualized equity across five company tiers at four career levels.

Sources.

Manually curated from Levels.fyi public role pages, the Carta H1 2025 State of Startup Compensation report, techinterview.org, and the Cadence engineering-comp blog, cross-checked against salary figures pulled from a 67,000+ job-posting corpus for the archetypes where postings were plentiful enough to yield a usable sample. This is not a live feed — treat it as directional, and expect drift as the market moves. Every cell carries a citation — including an explicit record for cells with no public data — and it's permanently archived, so none of it depends on a live URL that may go stale.

What the percentiles mean. Each band (P10 / P25 / P50 / P75 / P90) reflects individual-contributor total-comp submissions for the US market. The P50 is the median; the P25–P75 interquartile band is the thick middle of each bar. Equity is annualized paper value at vest, not liquid cash — and for growth-stage and early-stage tiers that equity is illiquid, so the equity figure is an expected value against a wide, uncertain outcome, not money in hand. Below is a real worked example for one archetype, Product / Full-Stack Software Engineer, so you can see what the P10–P90 track, the P25–P75 band, and the P50 tick look like in practice.

AI labs
$440k
FAANG / Mag7
$293k
High-growth public
$257k
Growth-stage private
$240k
Early-stage
$208k

product-full-stack-software-engineer · total comp (base + bonus + annualized equity) · P25–P75 band, P50 median

Confidence. Each cell is tagged high, medium, or low confidence by how much public data backs that specific role/tier/level combination. Cells marked low carry a limited data badge on the chart. 50 of 360 cells are low-confidence — concentrated in Customer Support Solutions Engineer (13 cells), Customer Support Engineer (10), Embedded/IoT Engineer (8), Developer Relations & Advocacy (6), Consulting Engineer / Professional Services (4), Sales Engineer (Pre-Sales) (3), Solutions Architect (Consulting) (3), and Engineering Management's L3 row (3, across tiers) — because no public compensation data exists for that specific role/tier/level combination. Engineering Management's L3 cells reflect a taxonomy mismatch on top of that: “L3 engineering manager” isn't a real leveling tier — first-line EMs are promoted senior/staff ICs, not L3 hires.

Not financial advice.

This is directional context to help you weigh paths, not a negotiation figure, an offer prediction, or personalized financial guidance. Your actual comp depends on company, location, level calibration, timing, and negotiation — none of which this dataset knows about you.

6. What's next

A future version layers in verified practitioners and hiring managers rating what their role actually demands day-to-day, with expert-seeded weights adjusted by that crowd input — but only once there's enough traffic per archetype to make that signal reliable, and always with the diff between expert and crowd published, not hidden.

Questions or feedback: kazaam@swe-genie.com.