Autonomous AI Customer Service Agent & Knowledge Engine
Stateful customer support agent built on LangGraph and FastAPI, running local Ollama inference and semantic retrieval over PDF, DOCX, and Excel manuals with zero cloud data transit.
The challenge
Automate customer support and ticket routing for enterprise products without leaking proprietary documents to external LLM APIs, while handling multi-turn context, human handoff, and file attachments.
Architecture & solution
Architected a LangGraph cyclic agent state machine on Python 3.11 / FastAPI. Extracted and chunked domain manuals across PDF, Word, and Excel formats into pgvector. Grounded local inference using Ollama (Qwen 2.5) with asynchronous tool calling for ticketing. Secured sensitive credentials using Fernet encryption and containerized the API, database, and Vite admin console with Docker Compose.
Architecture pipeline & data flow
Multi-turn support requests, customer tickets & multi-format manuals
Async REST endpoints, JWT validation & Fernet secret decryption
Stateful agent graph orchestrating conversation memory & tool calls
Multi-format parser (PDF/DOCX/XLSX) & pgvector semantic retrieval
Local Ollama (Qwen 2.5) private inference with zero cloud egress
Vite + Tailwind management console with Docker Compose volume isolation
Technical highlights
- Stateful multi-actor dialogue graph with LangGraph & LangChain
- On-premise private inference (Qwen 2.5) with pgvector hybrid retrieval
- Multi-format parser for operational manuals (PDF, DOCX, XLSX)
- At-rest credential encryption with Fernet & JWT-based session security
- Vite + Tailwind admin console with Docker Compose volume isolation
Tech stack
- LangGraph
- FastAPI
- Python 3.11
- pgvector
- Ollama
- Docker Compose
- Vite
- Tailwind