多代理 AI 客服平台

這是一個飯店業客服 AI 平台，我從 MVP 原型獨立打造到上線的多租戶 SaaS 產品：核心是一套多代理 LLM pipeline （意圖分類 → 身分解析 → 任務編排 → 訂房/RAG 子代理），透過聊天介面服務房客；後端與地端基礎設施同樣由我全棧負責。

✅ 從寫死的單租戶原型，重構為動態多租戶 SaaS，新租戶開通免改碼
✅ 設計分層、類 Mixture-of-Experts 的模型路由策略，每則訊息 LLM 成本降低約 60%
✅ 將服務從雲端遷移到地端自架，建立多環境部署流程並完成容量壓測
✅ 交付平台首位正式付費客戶

A hospitality customer-service AI platform I took independently from MVP prototype to a live multi-tenant SaaS product: the core is a multi-agent LLM pipeline (intent classification → identity resolution → task orchestration → booking/RAG sub-agents) serving hotel guests through chat, with the backend and on-prem infrastructure also owned end to end.

✓ Re-architected a hardcoded single-tenant prototype into a dynamic multi-tenant SaaS platform, enabling new-tenant onboarding with no code changes
✓ Designed a tiered, Mixture-of-Experts-style model-routing strategy, cutting per-message LLM cost by roughly 60%
✓ Migrated services from cloud to self-hosted on-prem, built a multi-environment deploy pipeline, and load-tested for capacity
✓ Delivered the platform's first live paying customer

OpenAI GPT-4.x · Multi-Agent Orchestration · RAG (pgvector) · Laravel / PHP · PostgreSQL · n8n · Docker · Nginx