NajNajNovije
A news aggregator that classifies coverage, not publishers.
- Type
- Product
- Industry
- News / media
- Platform
- Web application
- Role
- Design, build, deploy
- Status
- Live today: Croatian news, summarised and explained by AI, with an assistant for follow-up questions
What made this hard.
A Croatian news aggregator that pulls stories from many portals, clusters related coverage, and adds AI commentary and a political-spectrum reading, with every story linking back to its source.
Aggregating many sources means handling different formats, update cadences and duplicates, then staying fast when a breaking story spikes traffic.
The harder question was fairness. Labelling an article by the leaning of its publisher is easy and wrong: it mislabels neutral coverage from outlets with a reputation, and it is exactly the kind of shortcut a reader will catch.
What we built.
- Replaced source-level labelling with per-article classification: an LLM places each item in a cluster on a five-point scale, aggregated with weighted largest-remainder normalization. Shipped forward-only, with the source-based job retained as a documented fallback and rollback path.
- Added a regression-guard script that asserts named feature markers (schedule entries, command presence, front-end shape handling) so a feature can't disappear silently in a refactor.
- Built Croatian read-aloud on the browser Web Speech API: zero server cost, no keys, and engineering around real browser bugs: a heartbeat plus watchdog for mid-utterance stalls, sentence-boundary chunking for long-utterance drops, and graceful degradation when no Croatian voice is installed.
- Added a real-estate section with multi-source headless scraping, candidate-URL validation before insert (status, content and domain markers) so dead or fabricated links never land, and price-outlier flagging with shareable query-string filter state.
- Fixed the stats endpoint properly: composite indexes instead of full scans, and a “count not-null” rewritten as “total minus null” so an index actually applies. Also root-caused a JSON-path filter bug that extracted an object where a scalar was expected and silently coerced it to zero.
What it runs on.
Chosen to be maintainable for years, not to win an argument on the internet.
- PHP 8.2+
- Laravel 12
- Vue 3
- Vite
- Tailwind CSS 4
- MySQL
- SQLite (tests)
- LLM API
- Web Speech API
Where it landed.
Tests stand the news schema up on SQLite while keeping full-text index migrations conditional on MySQL, and fake the queue because creating an item dispatches enrichment jobs, which keeps the suite fast without pretending the pipeline doesn't exist.
Immutable cache headers on hashed build assets and longer cache TTLs on the hot endpoints keep the front page cheap to serve even when a story breaks.
Need something in this territory? This is Custom web app development work. Start from all services or see pricing.
Related case studies.
Your project
Tell us what you're building. We reply within two business days.