Building a Multilingual AI Call Center on a 4GB VPS
The article describes how the author built a production-ready AI-powered call center system using Django, Neo4j, Twilio, and other technologies on a $8/month VPS with 4GB RAM.
Why it matters
This project demonstrates how to build a robust, production-ready AI call center system on a low-cost infrastructure, leveraging various technologies and techniques.
Key Points
- 1Hybrid retrieval approach combining vector search and knowledge graph
- 2Multilingual pipeline that generates responses in the target language
- 3Separate AI experiences for customers and the business owner
- 4Efficient resource utilization on a low-cost VPS
Details
The author spent 18 months building RagLeap, a call center system that combines vector search and knowledge graphs for high-accuracy retrieval. The backend is built on Django, PostgreSQL, Neo4j, Celery, and Redis, running on a 4GB VPS. The system supports voice, WhatsApp, and Telegram channels, with Twilio and ElevenLabs for voice and messaging. It uses any AI provider (OpenAI, Anthropic, etc.) for language understanding. The key innovation is the hybrid retrieval approach, which blends vector search (75%) and graph-based search (25%) to achieve 94.3% accuracy, much higher than vector search alone. The system also supports multilingual responses, automatically generating answers in the target language set by the workspace owner. Additionally, the system has a separate AI-powered manager experience for the business owner, with over 50 executive actions registered.
No comments yet
Be the first to comment