SaaS platform for agentic AI customer service

Problem

Customer requests, whether a general inquiry, a sales matter, or an issue, need to be answered quickly, around the clock, and at consistent quality, including independent appointment booking, without a staff member having to step in every time.

Solution

For leezy.ai I designed and built a SaaS platform for agentic customer service: an LLM-based agent answers customer inquiries in real time, classifies incoming requests automatically (general inquiry, sales, or issue), and derives the appropriate action from that classification, up to booking appointments on its own. The platform has been in production since early 2024. A RAG architecture with a vector database runs in the background. The architecture was consolidated iteratively: from a Python-based prototype through a NestJS backend to an integrated Nuxt/Nitro application. I continuously measured answer quality and response times through Langfuse traces and optimized them iteratively.

Result

Result: around 85% of incoming requests are resolved fully automatically, without a staff member stepping in. The first response arrives in under 2 seconds, around the clock and at consistent quality.

Technical implementation

• Platform architecture, RAG pipelines, and prompt engineering strategies
• Multi-tenant architecture: each customer with its own knowledge base, its own integrations, and isolated data
• Automatic classification of incoming requests and independent selection of the resulting action
• Calendar integration with Google and Microsoft for automatic appointment booking
• Integrations with Zapier, Notion, Slack, and Microsoft 365
• Customer dashboard for reviewing requests and statistics
• Evaluation and prototype integration of LangChain and other frameworks for orchestrating LLM workflows