Building a decision engine for technology career guidance.
Years ago, a career quiz lived on a personal portfolio site — thirteen questions, twenty tech roles, and a scoring engine built from real mentoring experience. The site is gone. The engine wasn't. This is the record of finding it, understanding it, and rebuilding it into something wider and sharper for WEBMUSE.
A career quiz is easy to build badly and quietly useless.
Most "which career suits you" quizzes are a handful of generic questions mapped to a handful of generic outcomes — Product Manager or Developer, take your pick. That's not guidance, it's a coin flip with extra steps. The bar for a quiz worth someone's five minutes is higher: results specific enough to act on, and honest enough to admit when a path is a stretch rather than a lock.
Twenty-odd generic tech buckets isn't enough anymore — roles like platform engineering, SRE, and AI/prompt engineering barely existed as job titles when the original quiz was written.
A single percentage match tells you nothing about why. Two people can land on the same top match for completely different reasons.
The original quiz was built from real mentoring experience, not a stock personality-test template. Losing that voice in a rebuild would have been losing the actual value.
Rebuild the proven engine before you touch it.
The original site was offline with no local backup, but its source code was still public on GitHub — three successive quiz implementations inside one repository, each an iteration on the last. Before designing anything new, the actual scoring logic got read line by line. That reading shaped the whole rebuild.
Recover, don't guess
The real question set and scoring function were pulled from the original repository, not reconstructed from memory or invented fresh.
Expand the taxonomy
Twenty career paths became twenty-six, adding the roles that have emerged or matured since the original was written.
Add a second lens
A four-axis working-style score layered on top of the proven point-sum model, explaining the fit instead of just stating it.
Recovering a product that no longer exists.
A local filesystem search across the whole machine turned up nothing — no folder, no zip, no old build artifact. The one lead that actually worked was the developer's own GitHub account, where the original portfolio repository was still public. Cloning it surfaced the real question bank, the real career-path data (including salary ranges, roadmaps, and first-person advice), and the real scoring function — not an approximation of them.
The site was gone, the code wasn't
The live site at the original domain was down and no local backup existed. The actual source was still reachable as a public GitHub repository — three separate quiz implementations existed inside it, evidence of iteration over time.
Reverse-engineer the real scoring logic
Rather than guess at how the quiz worked, the actual scoring function was read directly: a flat per-question point-sum across 13 questions, normalized to a 15–95% display range, sorted to a top-3.
Don't discard a working model
The point-sum logic was preserved as the primary matching engine — it produces sane, explainable results. Nothing about it was broken; the job was expansion, not replacement.
A second scoring layer, not a bigger list
A four-axis working-style score (builder, analyst, connector, craftsperson) was added on top, so a result explains *why* a path fits, not only *that* it does — without touching the proven top-3 logic underneath it.
A scoring engine with two layers instead of one.
The original engine scored every answer directly against career-path IDs — pick an option, add its weighted points to however many paths it touches, sum across all questions, sort, take the top three. It's simple and it works, so it stayed. What changed is what runs alongside it.
| What changed | Detail |
|---|---|
| 13 → 14 questions | One new question added specifically to probe reaction to generative AI — the original set predates the current AI hiring wave. |
| 20 → 26 career paths | Added platform engineer, SRE, data engineer, AI/prompt systems engineer, technical program manager, product marketing, developer advocate, tech ethicist, open source maintainer, technical founder. |
| Flat point-sum | Original model: each answer adds weighted points directly to career-path totals, then sorts by total. Kept, because it works. |
| + Two-axis layer | New model: every answer also scores against four working-style axes (builder / analyst / connector / craftsperson), shown independently of the top-3 matches. |
Every quiz option now carries two payloads instead of one: the original scores map into career-path points, and a new axes map into the four working-style dimensions. Both are computed in the same pass over the same answers — the axis layer is additive, not a second quiz.
What didn't change on purpose.
Every career path in the original data carried a short, first-person "advice from Motun" line — not templated career-site copy, but the actual voice of someone who'd mentored people into these roles. That voice was preserved and extended to every new path added, rather than replaced with generic guidance once the roster grew.
The point-sum scoring model, the personal-advice format per path, and the zero-signup, instant-results experience.
The UI — from a shadcn/Radix component kit into WEBMUSE's own glassmorphism, electric-blue, Outfit/JetBrains-Mono design system.
6 new career paths, 1 new question targeting reactions to generative AI, and the full four-axis working-style layer.
Why these choices hold together.
Expand the taxonomy. Don't replace the judgment.
A bigger list of career paths is only useful if the matching underneath it is still trustworthy. The scoring logic that shipped originally wasn't broken, so it wasn't rewritten — it was read, understood, and built on top of.
The real question set and scoring model were pulled from source control, not reconstructed from a vague memory of what the site used to say.
A percentage match plus a working-style breakdown explains a result instead of just asserting it.
First-person advice per career path was preserved and extended — genuinely mentoring, not templated career-site copy.
No sign-up gate before seeing a result — instant answers were kept, matching what modern users actually expect.