Answers are grounded in this paper and cite their sources.
Traditional enterprise customer support systems rely heavily on rigid rule‑based chatbots or single large language models (LLMs). These approaches suffer from poor task delegation, hallucinated responses, and security risks when directly connected to business databases. This project addresses these limitations by designing and implementing a multi‑agent architecture that orchestrates specialised AI agents to handle customer inquiries securely and reliably. The work focuses on automating the “first line” of customer support for small businesses, including pre‑sales questions, order tracking, policy inquiries, and complaint resolution. The scope covers an end‑to‑end system comprising a customer‑facing storefront (EatIron/StayNest), a reusable chat widget, a FastAPI backend with DSPy‑based hierarchical multi‑agent system (6 agents), and a Ticket Management System (TMS) built with Tauri and React. The methodology uses DSPy for declarative agent signatures (Router, Catalog, DB, Policy, Resolution, Ticketing), pgvector for RAG, OpenAI GPT‑4o‑mini for routing/catalog and GPT‑4o for resolution/vision, and socket. IO for real‑time communication. Read‑only database access is enforced through a dedicated DB Agent to prevent data leaks. Experimental evaluation shows that the multi‑agent architecture achieves 92.4% accuracy after MIPROv2 optimisation (from 64.8% zero‑shot), correctly classifies intents, raises tickets only for genuine issues, and reduces owner intervention time by ~60% in simulated scenarios. The system successfully handles sales, support, and escalation flows with full transparency via a live dashboard. The architecture proves scalable and can be extended beyond customer support to HR, IT, and other enterprise domains.
No takes yet. Share an insight, caveat, or question.
Akash Anand (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: