← All work
SAASOperating2025
BeauBot
A LINE AI customer-service SaaS for beauty studios. Multi-tenant ReAct agent handles inquiries, bookings, reschedules. 5-step setup wizard — owners are live in under 10 minutes.
- Role
- Full-stack + product design
- Duration
- 2025 — present
- Reading
- 2 min read
Stack
GoEinoNext.js 16shadcn/uiPostgreSQL 18LINE Messaging APIOpenAI
Product Shots
SCREENSHOTS · COMING SOON
Screenshots coming soon. Drop captures into public/works/<slug>/ and add a screenshots array to the mdx frontmatter.
public/works/beaubot/01.pngTaiwan beauty studios (nails, hair, lashes) run on tight headcount. Owners are still answering "any slots tomorrow?" on LINE after closing — a real, painful problem. BeauBot turns that customer-service layer into an AI agent. Owners configure hours and services once, and AI handles bookings, rescheduling, and human handoff.
Problem
- Owners burn out from after-hours messaging
- Hiring a dedicated CS person costs more than studios make
- Existing LINE bots use canned templates — they can't handle "do you have a 3pm slot?" which needs comprehension and a database lookup
Solution
A multi-tenant ReAct agent + setup wizard:
- 5-step setup wizard: Studios configure LINE Channel, hours, services, staff schedule, and AI persona — averaging 10 minutes start-to-finish.
- Eino ReAct agent: Built on Cloudwego Eino. The AI is given 5 tools (lookup services, lookup availability, create booking, cancel booking, hand off to human) and plans steps autonomously.
- Tenant isolation: Per-studio LINE Channel, prompts, hours — same codebase, fully isolated data.
- Smart model routing: Trivial intents go to GPT-4o-mini; complex reasoning goes to GPT-4o. 60% cost reduction.
Stack
- Backend: Go 1.25 + Gin + GORM + PostgreSQL 18, Cloudwego Eino ReAct framework
- Frontend: Next.js 16 + React 19, shadcn/ui, Zustand, Tailwind CSS 4
- AI: OpenAI API (dynamic GPT-4o / 4o-mini routing)
- Integration: LINE Messaging API webhook + LIFF
- Deploy: Docker + Vercel + self-hosted PG
Outcomes
- Live with landing page (beaubot.ailoop.uk) and full admin
- Multiple test studios in production, AI auto-resolution rate > 70% (rest handed to human)
- Average response time < 2 seconds (LLM call included)
- Time-to-first-AI-booking after signup: under 30 minutes
Lessons
- ReAct isn't a silver bullet: Deterministic SQL beats LLM reasoning by 10x for fixed flows like "check availability". Agents win where conversation is fuzzy.
- Multi-tenant is more than schema: Prompts, tool permissions, API rate limits — all per-tenant. Otherwise one bad studio throttles everyone.
- Setup wizard is the prerequisite for paying: If they can't set it up in 10 minutes, they won't pay. More important than feature polish.