CoasterRank is a live community ranking product for roller coaster riders, designed, built, and run solo on free tiers.

Coaster enthusiasts have strong opinions about rides, but there were issues with the existing rankings out there on the internet. Either they were infrequently updated or methodology was suspect. Riders like to rank their favorites and see how their list compares to everybody else.

Constraints

- Solo project, zero budget; everything had to run within free tiers for hosting, data, and compute - Ranking quality first and transparency first - Copyright respecting catalog sourcing; growth through community input - Reproducible work, so the project doubles as a testbed for an agent-assisted engineering workflow - Privacy-respecting: no user tracking, no third-party cookies

Technical decisions

- Postgres for both data and compute: schema, ranking computation, and scheduled jobs - Bradley-Terry, computed in Postgres - Catalog with provenance: seeded from a CC0 dataset with 1,087 coasters, then grown through community submissions - Edge caching via Cloudflare for public reads - Agent-assisted workflow with guardrails

What I owned end to end

- Product design & scope: what a ranking app does and doesn't need - Data model & ranking pipeline, including the incremental Bradley-Terry fit and scheduling - Frontend in React and TypeScript: drag-sort rankings with optimistic updates, URL-mirrored search and filter, submission queue, and admin dashboard - Catalog seeding, submission review, and data quality - Growth loop: community posts, enthusiast outreach, SEO; signups, weekly active rankers, and submissions tracked in analytics and reviewed weekly - Operations: staying inside free tiers, cost checks, and the decision log

Stack

- Frontend: React, TypeScript - Backend: Supabase/Postgres, Cloudflare Workers, Edge Functions

Result

- Live product, approaching the first 100 users, running within free tiers - Over 100 community submissions so far - Open source: MIT code, CC BY user data - Continual optimization of the fit and storage to stay inside free-tier limits

What's next

- Hopefully, a few hundred or thousand more users! 🤩 😆 - Evaluating Plackett-Luce to replace pairwise Bradley-Terry and remove the O(n²) pair layer for disk space economy in DB