AI-driven news aggregator serving 20 countries
The regional press landscape is fragmented, and a multilingual audience couldn't be served by manual moderation. We built a fully automated AI pipeline — today it runs live across 20 countries.
- Ongoing product · 2025–
- Elvin Sharifov + 2 contractors (AI, infrastructure)
- Product design · Engineering · AI integration
- Live — 20 countries
Starting point
The regional information landscape is fragmented: dozens of sources per country, different languages, different editorial styles. Giving an audience a complete picture by leaning on manual moderation is both expensive and impossible to scale. Classic RSS aggregators were a stopgap — no quality filter, no multilingual distribution.
The brief was simple: a 24/7, multilingual, AI-driven news platform. Manual intervention only in edge cases — the system must run itself.
Decisions
- Split the pipeline into modules — source extraction, normalisation, classification, dedup, translation, distribution. Each module can stop and restart independently.
- Model selection by cost — Haiku 4.5 for fast extraction, Sonnet 4.6 for core work, Opus 4.7 for hard editorial calls. Prompt cache drops cost by 60%+.
- Human review only on anomaly — the system flags low-confidence items for a person, not every article.
- Multilingual distribution — the same story goes live in 3–4 languages in parallel.
Execution
Stack: TypeScript + Node.js + Postgres + Cloudflare Workers. AI: Anthropic SDK plus local embedding models. Aggregation runs on Cloudflare cron + queues; distribution lives on a separate edge tier. Sentry for backend monitoring, an internal dashboard for visibility.
“The hardest part wasn't model selection — it was dedup logic. The same story arriving from two sources in different wording is normal. Embedding cosine similarity cracked it for us.”
— Elvin Sharifov
Result
- Live in 20 countries — audiences get a global picture from one source.
- 100% automation — manual moderation only on anomaly.
- 1000+ sources aggregated, deduplicated, and distributed across languages.
- The internal monitoring dashboard was handed over in Phase 4 — the team now runs the pipeline themselves.