Aurore is a skincare notebook. Track products you own, want, or ruled out: INCI list, dated notes, the reason behind each decision. The point: never redo the same research twice.
Not a medical tool. No score, no verdict, never tells anyone what to buy.
Live demo → aurore-app.fr. Open beta, French interface. One click creates a demo account: no email, no form, a collection already filled in.
browse the catalogue → read the formula → put the product on your shelf
→ write what happened → decide, and keep the reason
Five shelf states, because "no" is information too:
| State | Meaning |
|---|---|
| En stock | You own it, it is in use |
| Wishlist | Not bought yet, you intend to |
| Garde un œil | Bookmarked, no commitment |
| Archivé | Finished, or set aside |
| À éviter | Rejected. The note says why, so the research is not repeated |
Catalogue: public, no account. 7 300 products from 267 brands: 5 500 skincare, the rest hair, oral care and supplements, with INCI lists parsed into 1 000 linked ingredients. A product page reads the formula ingredient by ingredient, with tags, brand and discussion threads; an ingredient page gives its role, its family and the products that contain it. Search and filters cover the whole catalogue. Anyone signed in can submit a missing product or ingredient and suggest edits; a moderation queue reviews them.
Your shelf: with an account.
- Dated notes and a quick sentiment on each product. "Saint Graal" is the top sentiment, not a sixth state.
- Purchases: price paid, opening and finishing dates, expiry. What a product actually costs you.
- Comparison: 2 to 8 products side by side, formula against formula, saved and reopenable.
- Motifs: what the formulas already on your shelf have in common, as recurring signals. Never a ranking.
- Skin profile: goals (redness, hyperpigmentation, ageing, acne) and sensitivity. It changes what is highlighted while reading a formula. A declared preference ("I avoid X") can hide a product, but the masking stays visible and reversible in one tap.
- Public profile page and a feed, for people who want to share their shelf.
Around it: blog written in-app by the maintainers, full data export, account deletion, no third-party tracker.
Derived from each product's INCI list: possible benefits, points worth knowing, and a confidence level that drops when ingredients are unknown. Not a diagnosis, not a recommendation, not a safety score.
The logic lives in algo-derm, a separate MIT library vendored as a backend
tarball. The browser only ever receives the final ProductAssessment, never the ingredient
dataset. Contract, confidence model and limits: docs/scoring.md.
| Catalogue | Formula reading | Ingredient reference |
|---|---|---|
 |
 |
 |
Full-stack TypeScript on Bun, end-to-end typed without codegen: the backend exports its AppType,
the frontend consumes it through the Hono RPC client.
Bun · Hono (REST + typesafe RPC) · React 19 / TanStack Start (SSR via Nitro), Router, Query · PostgreSQL 18 + Drizzle ORM + Row-Level Security · Zod schemas shared by both sides · Vanilla CSS, Lucide icons · Biome, bun:test, Vitest, Playwright, Lefthook · Docker Compose, Nginx.
aurore/
├── backend/ # Hono API: routes, services, database access
├── frontend/ # React app, Vite + TanStack
├── shared/ # Zod schemas shared by both
├── vendor/ # Vendored dependencies, including algo-derm
├── infra/ # Docker, Nginx, ops config
├── backups/ # Database backup workflow
├── scripts/ # Automation scripts and just recipes
└── docs/ # Project documentation
Auth is email/password with Argon2, plus Google OAuth: a short-lived access token and a rotating
refresh token in an HttpOnly cookie. User-owned rows are isolated by PostgreSQL RLS, not by
application code. Details in docs/SECURITY.md.
Needs Bun 1.3.12, Docker with Compose v2,
just 1.51. Optional: mise installs the pinned
versions from .mise.toml and the git hooks.
just init # dependencies, generated JWT secrets, .env.dev, hooks when mise is available
just dev-fresh # typecheck, build and start the stackjust init writes a .env.dev that boots as-is: third-party keys (Brevo, Google OAuth, Bunny
CDN) stay placeholders, not needed to browse the app. On first run, answer yes when offered the
committed catalogue snapshot. Declining leaves an empty database and an app that errors on every
page.
Open http://localhost:5173 and click « Créer un compte de démo ».
Daily work is two terminals: just ts-check (host TypeScript watch) and just dev (Docker: runs
a host TypeScript preflight first, since the containers run the sources, not the build).
Some checks run with Bun alone, no Docker, no database, no E2E:
bun install
(cd shared && bun run build && bun test)
(cd frontend && bun run test:run)
bunx biome check .docs/commands/: every runnable command, grouped by task. Its README is the index, and it covers the recipes that write as well as the read-only ones.docs/conventions/: the cross-package conventions, one page per topic.docs/scoring.md: formula reading contract, confidence model, limits.docs/adr/: architecture decision records, numbered. Its README is the index.docs/SECURITY.md·docs/PRIVACY.md: auth model, role separation, GDPR handling.
Not dermatological advice, not a diagnosis. Formula reading depends on ingredient coverage:
unknown ingredients lower confidence, and a low-confidence reading is not a signal. Reading is
per product, routine-level interactions aren't modelled. Interface is French only. algo-derm is
pre-1.0, its calibration still moves. Open beta, one developer, tested at personal scale.