Data-driven npm package comparisons. Real stats, no guesswork.
DevPick helps developers choose between npm packages by providing side-by-side comparisons with real data: weekly downloads, bundle sizes, GitHub activity, maintenance scores, and AI-generated verdicts with 12-16 weighted criteria per comparison.
| Metric | Value |
|---|---|
| Tracked packages | 205 |
| AI-generated comparisons | 478 |
| Categories | 45 |
| Static pages | 733 |
| Build time | ~7s (with OG cache) |
| Full AI regeneration cost | ~$0.12 |
| Layer | Technology |
|---|---|
| Framework | Astro 5 + React 19 (Islands) |
| Styling | Tailwind CSS v4 |
| Database | Neon PostgreSQL (serverless) |
| ORM | Drizzle ORM |
| AI (content) | Gemini 2.5 Flash Lite (via OpenRouter) |
| AI (evaluation) | GPT-4.1 (via OpenRouter) |
| Search | Fuse.js (client-side) |
| OG Images | Satori + resvg-js |
| Hosting | Cloudflare Pages |
| CI/CD | GitHub Actions |
Nightly Pipeline (03:00 UTC daily)
flowchart TD
A[npm Registry + GitHub API + bundlejs/esbuild] --> B[(Neon PostgreSQL)]
B --> C[Gemini 2.5 Flash Lite<br/>metric validation + retry]
C --> D[Astro Static Build<br/>733 pages in ~7s]
D --> E[Cloudflare Pages]
E --> F[IndexNow API]
Weekly Discovery (Monday 06:00 UTC)
flowchart TD
A[npm search<br/>38 categories x 2 terms] --> B[10-layer filter<br/>deprecated / low downloads / vendor SDK]
B --> C[GPT-4.1 candidate scoring<br/>0-100, min 70 to pass]
C --> D[Auto-approve max 3/week]
D --> E[git commit --> nightly picks up]
- Fully autonomous growth: Weekly pipeline discovers, evaluates (GPT-4.1), and approves new packages without human intervention. The site grows itself.
- Multi-tier AI routing: High-profile packages (React, Vue, Tailwind) get enriched context injection. Evaluation uses a separate, more expensive model than content generation.
- 4-tier bundle size fallback: bundlejs API → bundlephobia → local esbuild (browser) → local esbuild (node) → node_modules install size. Never fails silently.
- AI metric validation: Regex-based fact-checking layer catches hallucinated stats in AI output. Incorrect claims trigger a correction retry before saving.
- Content-hash OG cache: Satori-generated PNG images are cached by content hash. Build time dropped from ~227s to ~7s.
- Budget guards: $5/day for nightly pipeline, $1/day for weekly discovery.
BudgetExceededErrorhalts gracefully. - Build-time query memoization:
once()wrapper reduces 77 DB queries to ~12 across 733 pages.
Source code is in a private repository. This repo serves as a public project overview.
Mehmet Aras — Senior Frontend Engineer