Astrolab

A free, privacy-first career-guidance platform for teenagers and adults. Live at astrolab.nikam.dev — no registration, no paywall, no course to buy.

Astrolab helps a person at a crossroads see a map, not a verdict. In about 12 minutes you take a short assessment based on the RIASEC model (Holland Codes), get a profile of your interests and work values, a fan of fitting occupations with an honest "why this suits you", and — the part most tools skip — the concrete path to get there: fields of study, the exact exams to take (Russian ЕГЭ first), and application deadlines.

It was built for a real person (a nephew choosing where to apply after school) and then opened to everyone, because the same problem hits hundreds of thousands of families every year. It is a social project, not a business — monetization is designed into the architecture but switched off.

---

The problem it solves

Career guidance online is broken in a few places at once:

• The test is bait. Most "free" career tests are lead magnets for paid consultations and courses; the result is deliberately cut short so you reach a payment button.
• Guessing instead of a method. A lot of them are "which fruit are you" quizzes with no model behind them.
• A test, but no path. Even a decent test usually ends at a list of professions. A person needs the next link: which field to study, what to take, when to apply.
• Adults barely exist. Millions want to change careers at 30–45, but almost all content targets schoolchildren.

Astrolab closes exactly these gaps.

How the assessment works

Not "one more quiz" — a chain from "I don't know who I am" to "I know where to apply."

1. Assessment (~12 min). Short cards with no right answers, across a few blocks:

   Block	Draws on	Format
   Interests	RIASEC / Holland (international standard; O*NET tags every occupation with it)	swipe cards, like / meh / no
   Cross-frame	Klimov ДДО typology, mapped from RIASEC	second lens on the same answers
   School subjects	like × good-at per subject	affinity grid
   Work values	O*NET-inspired work values	forced-choice pairs — "which matters more, A or B", so real priorities surface instead of socially-desirable ones
   Traits (optional)	TIPI-10 short Big Five	opt-in
   AI interview (optional)	adaptive dialogue	surfaces contradictions and hidden interests; fully degradable

2. Profile. Your interests across the six RIASEC types and what matters to you at work — shown visually, not as a dry list.

3. Occupations, in three baskets. Core (strong matches), Nearby (worth a look), and Dark horses (unexpected, to widen the field). Each occupation comes with a warm, honest explanation of why it fits you.

4. How to get in. Per occupation: fields of study (with specialty codes), typical exam combinations (ЕГЭ), a deadline calendar, and links to the universities and colleges that offer the field.

Methodology & honesty

This is the part we refuse to fudge.

• The match is a transparent formula, not a black-box model. Occupation fit is a deterministic score (cosine similarity over RIASEC and values vectors + subject affinity − soft penalties). The same answers always produce the same result, and the ranking logic is explainable. Weights are a documented prior, not a validity claim, and the scoring config is immutable so any past result is reproducible.
• AI is used in exactly one place. It takes the already-selected occupations and reorders / explains them for a specific person. The LLM never enters the score and never blocks a result. It does not decide who you should be and does not invent occupations.
• It's an assessment, not a diagnosis or a verdict. Your profile changes with experience — retake it every six months. We say "based on the RIASEC model" and deliberately do not use the word "validated" until we can demonstrate it on our own data. When a choice affects years of someone's life, being honest about the limits of the method matters more than a confident promise.

Privacy by design

Because many users are minors, privacy is in the foundation, not bolted on:

• Anonymous by default — the assessment runs with no registration.
• An account is only needed if you want to save a result and come back to it.
• Minimal data collection; tokens are hashed; deletion and retention are in from day one. We do not sell what we collect.

For adults — a first-class case, not an afterthought

Changing careers is a huge, underserved segment. An adult takes the same assessment, but instead of school subjects there's an experience step (paste a résumé or describe your background). Recommendations are then framed as a transition, not a restart from zero — the system points out what already transfers. "Eight years in sales" isn't "start over"; it's a concrete set of transferable skills that is worth a lot somewhere.

Why it's free

It's a social project. There are no paid tiers, courses, or "expert consultation" at the end. Charging a teenager to understand themselves feels wrong. Monetization exists in the architecture for the future but is turned off — the service runs free and without limits.

---

Tech

• Backend: FastAPI · SQLAlchemy 2 · Alembic · PostgreSQL.
• Frontend: SvelteKit (SSR, adapter-node) · Paraglide i18n (Russian-first, locale-agnostic — every UI string is a translatable key).
• AI: an optional, degradable layer (reranking + explanations only).
• Russian-first product; built so other locales are drop-in.

Repository layout

backend/    FastAPI + SQLAlchemy 2 + Alembic (schema + migrations)
frontend/   SvelteKit SSR + Paraglide i18n (/ru, /en)
docs/       METHOD.md · DATA_SOURCES.md · PRIVACY_MODEL.md · ARCHITECTURE.md
deploy/     systemd units, Caddy snippet, deploy webhook, provisioning
scripts/    check_cyrillic.py (CI guard: no hardcoded UI strings in source)
docs        internal design notes (kept private)

Local development

Backend (needs a local Postgres, or point ASTROLAB_DATABASE_URL at one):

cd backend
python -m venv .venv && . .venv/Scripts/activate   # or bin/activate
pip install -e ".[dev]"
alembic upgrade head
uvicorn app.main:app --reload --port 8015

Frontend:

cd frontend
npm install
npm run dev          # or: npm run build && npm run preview

Quality gates (CI)

• scripts/check_cyrillic.py — no hardcoded Cyrillic in source (keys only).
• Backend: ruff, mypy (non-blocking), alembic upgrade head from zero, pytest.
• Frontend: eslint, svelte-check, vitest, SSR build.

Deploy

git push origin main → webhook → the server pulls, migrates, rebuilds, and restarts astrolab-api + astrolab-web. Secrets live in /etc/astrolab/env, never in git.

Principles worth not breaking

• The LLM never enters the deterministic score, and never blocks a result.
• scoring_config is immutable (DB trigger); results are reproducible.
• Anonymous-first; tokens hashed; deletion + retention from day one.
• Nothing is published to the occupation catalog without human review.
• "Based on the RIASEC model" — never "validated" until calibrated (see docs/METHOD.md).

Docs

• docs/METHOD.md — assessment methodology, scoring, honesty rule.
• docs/DATA_SOURCES.md — occupation and admission data.
• docs/PRIVACY_MODEL.md — data, retention, minors.
• docs/ARCHITECTURE.md — services, deploy, infra.

Author

Built by Nikita Amosov — I make focused products that solve one real problem end to end, not platforms-for-everything. More at nikam.dev.

If Astrolab is useful, share it with someone who's deciding where to go next.