Your résumé says SQL.Can you show it?

from claim to evidence

Every line on your CV, offered up for testing

A career track is a job with its six topics named. Open one cold and it tells you where you stand immediately: nothing assessed yet, and here are the six things this role wants evidence of.

Paste your résumé in, or photograph it. SkillCompass reads it against that track's own curriculum and marks which of the six your experience already evidences. Then it does the part that matters: each evidenced topic gets a Verify button next to it, because a résumé claim is only a claim. Verifying means sitting the two-minute check and finding out whether the thing you wrote down survives six questions written by someone who wasn't trying to flatter you.

The text is used for one call and never stored, pasted or uploaded. It doesn't reach a database and the result stays on your device. A job-postings corpus would let this compare you against live market demand as well. There isn't one yet, so the report compares against the curriculum and says so on the page rather than inventing salary bands.

This gap report compares you against this track's own curriculum. A live job-market panel (top skills in this month's postings, salary context) arrives in a later beta phase.

the footnote under every gap report

two minutes

Six questions, pitched at you

One placement question comes first and sets the band. Answer it well and the six that follow come from the hard set; miss it and they come from an easier one, so the check spends its questions where they tell it something. You can skip the placement and pick the band yourself.

Nothing is timed. Your pace is recorded and handed back afterwards as seconds per question, because how long you took is information. It never costs you points.

where the questions come from

Every answer names the book it came out of

Answer and the explanation opens straight away, right or wrong. It says which option was correct, why the tempting one isn't, and then cites the public reference it was written against. Below, that's Davis and Goadrich's 2006 paper on precision-recall against ROC curves.

The citation is doing real work. A multiple-choice question with a debatable key is worse than no question at all, so each item is written against a named source, then run past an independent review pass that hunts for exactly the failure modes questions fail on: ambiguous keys, giveaway option lengths, trick phrasing. A person reads it before it counts. That is 645 questions so far, across 230 concepts.

what you walk away with

A map of the topic, not a mark out of six

The score is one line. Underneath sits the knowledge map: every concept the topic covers, coloured by how you did on the questions that touched it, with the prerequisite links drawn between them.

Concepts your six questions never reached stay dashed and empty. The map would rather show you a hole than guess. Tap any node for what the concept is, how to close it, and a way straight back into fresh questions on it.

A clean 6/6. You handled ROC-AUC vs PR Curve, the hardest idea in this set, without slipping. Take it again to cover confusion matrix; a perfect run on fresh questions is the real test.

the diagnosis written for that run

how difficulty stays honest

Boring arithmetic, deliberately

A question's difficulty starts as an estimate and then moves toward what people actually score on it. The step shrinks as an item collects answers, so a well-established question stops swinging on any one person's bad morning.

Ability starts at 1350, 1500 or 1650 depending on which way the placement question sent you. Only first exposures feed calibration, so retaking a topic can't inflate the statistics. All of it sits on a public page rather than in a footnote.

the number it won't invent

No percentile until 30 people have sat it

"Top 27% of 412 this month" means precisely that: your result placed against every real session on that topic in a rolling 90-day window. Until a topic clears 30 sessions there is no such comparison to make, so the result says early data and stops talking. Your first attempt is the one that fixes the percentile, so retakes draw unseen questions but can't be used to shop for a better rank.

The community dashboard runs on the same rule. It's built and wired, and for now it reports the only thing it can honestly report.

how much is actually live

385 topics mapped. 29 of them ready.

The catalog is the single source of truth: 385 subtopics across 46 fields, feeding 117 career tracks. Software, cloud, security, marketing, product, project management, HR, healthcare, law, economics, accounting, maths and psychology sit alongside the original data and finance core.

A subtopic goes live the moment its reviewed bank lands, and nothing is edited by hand to make that happen. The manifest works out live, partial or soon from whatever content exists, so topics, fields and tracks flip over by themselves. That's why the count above is a real number rather than a roadmap. Five fields are open now, four tracks are complete, and 42 more are partly covered.

the capstone

An interview that asks a second question

Get all six topics in a track assessed and the mock interview unlocks. Six to eight turns of real conversation, pitched at the seniority your results demonstrated rather than the one you claimed, and it follows up when an answer is thin instead of nodding along. It ends with a scorecard broken out by dimension. Tested against a deliberately vague answer, it flagged it and kept pushing, which was the point of the test.

It holds no server-side session. The transcript is resent each turn, so there is nothing sitting on a server to leak. The third feature is the study guide: a downloadable plan for a whole track, built from your real per-concept results and, where they exist, your résumé and interview too.

your data

Two buttons: export, and erase

Everything you do lives in one localStorage key in your own browser. The Progress page reads it back to you, every check with its date and score, and gives you a button to take it away as a file and a button to delete the lot. Neither asks you to sign in first, because there is nothing to sign in to.

The one thing recorded centrally is an anonymous answer log: question id, right or wrong, how long you took. That's what calibration runs on. It never records who you are, and there's no account for it to hang off.

smaller things worth knowing

Details that only matter once you're using it

what it runs on

Free, and free to run

SkillCompass costs nothing a month to operate. That's a constraint rather than a brag: nothing here depends on a budget that could get cut, and no part of it has to be monetised later to survive.

Frontend

React, Vite and TypeScript, served as a static build from GitHub Pages. The whole quiz loop runs here, so the core of the product can't go down with a backend

API

Python FastAPI on a Hugging Face Space, free CPU tier. Percentiles and the three AI features only

Calibration store

Supabase Postgres, free tier, reached through a small stdlib PostgREST client rather than an SDK

AI features

Gemini 2.5 Flash on the free tier over plain REST, with thinking tokens switched off so answers don't get truncated before they finish

Question pipeline

A generator drafts in difficulty bands, an independent reviewer pass repairs or rejects, and a person signs off before anything counts

Curricula

Referenced against Coursera, edX and Kaplan. Questions cite named public sources on the question itself

Your data

One localStorage key. An anonymous answer log is the only central record