How We Work

The full Hamidun News pipeline — from RSS collection to publication in 6 languages — and the quality checks at every step.

01

Pipeline

RSS+SPA collect every 30 min → Haiku pre-filter (4 parallel workers) → quality scoring → Haiku rewrite (short version 250–400 words, full 800–1,500 words) → Gemini image cover → 5-language translation → editorial publish slot → Telegram channel + portal + mini-app.

02

Quality controls

  • pre_filter_score — relevance plus topical quality (Haiku verdict).
  • quality_score — post-rewrite QC (transliteration regex, length check, retry-once-then-downrank).
  • is_duplicate_of — cross-language semantic dedup (named-entity overlap within a 6-hour window).
  • A 1-hour delay-after-approval before bot promotion — lets reader reactions surface.
03

Languages

  • Russian is the source of truth (humans + Haiku).
  • 5 translations (en/es/pt/fr/ar) via Gemini 2.5 Flash with retry-on-empty.
  • Untranslated articles get noindex,follow on that language and are dropped from hreflang clusters.
04

Algorithm transparency

  • The Telegram bot delivers the top-8 per 24h by quality_score (the BOT_TOP_N_PER_24H setting).
  • The portal homepage shows newest-first with category chips; signed-in users get a personalized “For you” rail (anonymous affinity by category clicks).
  • We do not boost paid content. We do not de-rank competitors.