Veloci

A multilingual AI news surface covering Telugu, Tamil, Kannada, Malayalam, Marathi, Hindi, Bengali. Each language gets its own retrieval, summarization, and rendering path. Shared orchestration only.

Status: prototype. Reader UI shipped; ingestion + summarization pipeline runs locally, not yet productionized.

• Reader — Next.js, TypeScript, Tailwind.
• Summarization — a multilingual LLM with per- language prompt templates.
• Embeddings — multilingual-e5 for cross- publisher dedup where it makes sense.
• Backend (planned) — worker service polling publishers, writing to Postgres + pgvector.

• Per-language evaluation data isn’t off-the- shelf. Building a small hand-labeled set per language is the bottleneck.
• Cross-publisher dedup is noisy. Same story, different wording. Source-citation per claim is the discipline that keeps it honest.
• Editorial voice differs per region. Translating-and-flattening produces translated-English. Per- language prompt templates preserve cadence.