raziel

Inspiration

Every week there's a new "best AI model" thread, and every week it's the same problem: a wall of leaderboards and marketing copy that answers a question nobody actually asked. What people usually need is simpler — "for this specific job, which model should I try?" — and almost nothing answers that honestly, in Spanish, without pretending to be a scientific ranking.

Raziel started from that gap: a human, curated, editorial catalog of AI models — think "IMDB + Rotten Tomatoes" for AI — Spanish-first, with English as a first-class second language.

What it does

• A searchable, filterable catalog of AI model families and variants.
• Model detail pages: ratings, benchmarks, architecture, license, local-execution taxonomy, release timeline.
• A comparator for 2–3 models, reproducible via URL parameters.
• /choose: a short recommendation guide that asks a few questions and applies transparent editorial rules over curated data — no AI black box, no fake confidence score, just visible reasoning.
• Radar: an editorial feed of model releases linked back to catalog entries.
• Full light/dark themes and es/en localization.

How we built it

Next.js 16 (App Router, Turbopack), React 19 Server Components, TypeScript strict mode, and Zod as the single data contract over versioned, hand-curated JSON — no backend, no scraping, no DB yet. We used a vertical-slice architecture (app → features → entities → shared), next-intl for i18n with all copy in messages/{es,en}.json, and a custom dependency-free theme store instead of next-themes. /choose maps user answers to a canonical use-case taxonomy and returns one main recommendation plus up to two alternatives — never padded. Data-quality tooling (data validation, dead-code analysis, rating recalibration) keeps the catalog honest as it grows.

Challenges we ran into

• Rebuilding /choose so it reads from the shared use-case taxonomy instead of a hand-written model-ID tree that broke every time the catalog changed.
• Being honest about sparse data — some use cases have only one evaluated candidate, so the flow shows up to two alternatives (sometimes zero) instead of padding with irrelevant models.
• Keeping the visual language editorial rather than generic-SaaS, after an internal audit flagged AI-slop drift — this led to a canonical design system and a unified background/surface hierarchy across all public routes.

Accomplishments that we're proud of

• Shipping a fully bilingual (es/en) editorial product — 23 model families, 47 variants — that is curated by hand and validated end-to-end with Zod, with zero backend infrastructure.
• Building /choose, a recommendation guide whose every decision is traceable to a visible rule and a reviewed data point, instead of an opaque AI verdict.
• Keeping the codebase honest as it grew: a strict vertical-slice architecture, automated dead-code and data-coverage checks, and a written decision log that makes the project's own history as inspectable as its code.
• A deliberate visual identity — light/dark themes, a unified editorial canvas — that reads as a curated product rather than a generic AI-tool template.

What we learned

That an AI catalog is more interesting when you refuse to let AI make the actual decisions — every /choose rule is legible and arguable, which builds more trust than a score ever could. And that keeping a data-first architecture honest takes real discipline: every shortcut (a hardcoded ID, a one-off mapping) is a future bug.

What's next for Raziel

• A lightweight database and private editorial CRM, migrating progressively off static JSON into live data.
• Deeper catalog enrichment with per-datapoint sourcing, and a real newsletter.
• Public auth

Built With

• javascript
• next-js
• react