AI · Automation
AI support assistant
An agent that closes 70% of tickets without a human. RAG over the knowledge base.
NDA project. An LLM + RAG support system: classification, answers to common questions, routing of complex ones.
Challenge
The client's support team was buried in routine questions - response times slipped and quality dropped. The goal was to cut manual load without losing control of the customer conversation.
Approach
- Connected an LLM to the client knowledge base via RAG for accurate answers to common queries.
- Designed the pipeline: classify → search the KB → generate an answer with a confidence score → escalate to a human.
- Built an agent interface: the system suggests an answer, the operator edits and sends it - human-in-the-loop.
- Added failure analysis so every Q&A pair feeds back and improves the system.
Stack
PythonFastAPILangChainLLM / RAGQdrantPostgreSQLTypeScript / ReactKubernetes
Highlights
- −70% manually handled tickets
- Common queries answered several times faster
- Control retained: a human reviews each answer before sending
- Self-improving on real conversations
Result: −70% load