Case study · 2026
Anime Compendium
A library of every anime I’ve watched, opened up to visitors: they can browse, react, comment, and see when the next English sub or dub episode lands in their own time zone.
Role
- Design, architecture, and build (solo)
Stack
- React
- TypeScript
- MUI
- React Query
- tRPC
- Fastify
- Prisma
- PostgreSQL
- Turnstile
- Playwright
The idea
I wanted one place for everything I’ve watched, organized the way I think about shows (shelves like “Chill/Atmospheric” and “Psychological/Mystery”), not the way tracking sites do.
It grew into a public site: visitors can browse the library, leave a ❤️ or 👎 on a show, comment, and check English release times.
English releases, in your time zone
Simulcast schedules are published in Japan time, and English sub and dub releases land at different times again. The Releases page shows the week’s English episodes converted to each visitor’s own time zone, and can narrow it to shows in the compendium.
A rating filter that matches what you see
Ratings are stored out of 10 but shown as half stars, so a “4 stars and up” slider quietly disagreed with the cards: a 9.4 looks like 4½ stars, yet number ranges sorted it inconsistently.
I replaced the slider with star chips whose ranges come from the same rounding the cards use, and a test covers every rating from 0 to 10 so the two can never drift apart.
Try it · sample data
- Vinland Saga★★★★½
- Cowboy Bebop★★★★½
- Spy x Family★★★★
Spam protection without tracking people
Reactions and comments are open to anonymous visitors, which invites bots. Cloudflare Turnstile blocks scripts, per-connection limits survive restarts, and an admin page flags unusual bursts by country and network type.
None of it stores an IP address: connections are grouped by a keyed hash that changes every day.
Comments go through a review queue before they appear, with spoiler tags and auto-generated display names so nobody has to invent a username.
Personality in the details
The site has a voice: comments are “Unsolicited Drivel”, the suggestion form asks you to “Convince me”, and for anyone overwhelmed by all the features there’s an “Overstimulated” page with one long, blissfully simple list.
Decisions and trade-offs
A hobby project, built like a product: each of these was a deliberate choice with a cost.
One typed contract from database to UI
tRPC infers the client’s types from the server’s router, so there’s no hand-written API client to drift out of date, and zod validates every input. The trade-off is tying the frontend to a TypeScript server, which is fine for one codebase. Login stays plain REST: two fields, nothing worth sharing.
Real migrations, guarded in CI
Schema changes are committed migrations. CI builds a fresh test database from them exactly as production does, fails if the schema drifts from them, and blocks any migration that could destroy data unless it’s explicitly marked as intended.
Privacy over precision
Spam protection uses Cloudflare Turnstile and per-connection daily caps keyed by a hash that changes every day, never a stored IP address. Unusual activity is flagged for review rather than blocked by country. I give up some precision in exchange for treating visitors the way I’d want to be treated.
A small model where a word list fails
“This deserved 5 stars” on a 3-star show is a disagreement even though it sounds positive. A small, cheap model (about $0.0001 per comment) reads each comment against my actual rating and review, and flags likely spoilers in the same call. A hard spend limit is the backstop.
Modules by domain
The server is split by what things are (entries, comments, releases, shelves), each owning its router, services, and data access. New features land in one folder instead of spreading across layers.
When it broke
Early on, a test run pointed at the real database instead of the test one, and its setup step deleted every entry in the library. I restored it from a backup; only a few hours of edits were lost.
Then I made it impossible to repeat. End-to-end tests now refuse to start unless the database name ends in “_test”. Backups run automatically, with a job that checks every hour, verifies each copy by counting rows, rotates old copies, and keeps one off the machine. Any risky command backs up first, and a written runbook covers restoring into a fresh database before swapping it in.
What I’d do differently
Set up automated backups and the test-database guard on day one, not after an incident. And design visitor privacy and spam protection before opening reactions to the public, rather than retrofitting them.