Emergency Alert System for 15,000 Users: LLM, PostGIS, Qdrant, and Telegram

The article details the architecture of an emergency alert system serving 15,000 users. The primary engineering challenge was automating the processing of unstructured data from hundreds of Telegram channels without manual moderation. To address this, the authors built an LLM-based pipeline that classifies events, identifies locations, generates summaries, and removes duplicates. The technology stack includes PostGIS for geospatial data and the Qdrant vector database for semantic search. The system also incorporates crowdsourcing, allowing users to report events and verify each other's messages. This approach created a scalable solution capable of operating 24/7, ensuring the rapid delivery of critical information to end users.