Hanami Hair Studio

Context

Hanami is a hair salon in the historic center of Zacatecas with a clear philosophy: natural, inclusive styling. Clients discover it on Instagram and arrive on their phones — first to look, then to ask by DM.

The problem

Most conversations started with the same questions — prices, how long a session takes, what's needed before a color service, the deposit — before reaching the one that matters: is there a slot?

A generic chatbot wasn't enough. Answers had to match the salon's real prices and policies, and availability had to match the owner's real calendar. In booking, a wrong answer is worse than no answer: a client who shows up for a slot that doesn't exist is a client you lose.

Architecture

The system has two surfaces with a clear split. The site answers the static questions (prices, policies, what to expect). The bot handles the dynamic ones (availability and booking). They meet in one place: every "book" button on the site opens an Instagram DM with the bot.

HanamiBot

Rendering diagram…

• Channel. Instagram DMs reach n8n through ManyChat, and replies go back the same way.
• Pre-processing. Voice notes are transcribed and images described with Gemini, and a 20-second buffer merges messages that arrive in a burst, so the agent answers the whole thought instead of every fragment.
• Classifier and router. A Gemini classifier labels each conversation's intent and a router hands it to one specialist agent.
• Booking agent. Books, reschedules and cancels in the owner's Google Calendar, and answers questions about services, prices and policies from a knowledge base in Postgres with pgvector.
• Color agent. Color services need a consultation, so this agent collects a five-point checklist about the client's hair, saves it and hands the conversation to the owner. It never quotes a final price.
• Free-slot calculation. The only source of availability. See the hard part for why that's code, not a prompt.
• Voice. A last step rewrites each reply in the salon's tone before it's sent.

The website

hanamihairstudio.com is a React and TypeScript single page built with Vite and Tailwind, prerendered to static HTML at build time, so the content is in the page before any JavaScript runs.

• Booking information up front. Prices, session length, opening hours, payment methods and the deposit policy for color services are on the page before anyone gets in touch. The goal: clients reach the DM already qualified, without repeating questions.
• Every call to action opens the bot. "Book" buttons are Instagram DM deep links, so the site hands off to the bot instead of to a contact form.
• Before/after gallery. A scroll-snap carousel on mobile and a grid on desktop, with its animation skipped for visitors who prefer reduced motion.
• Local SEO. Schema.org HairSalon structured data with the address, opening hours, price range and a catalog of services.

QA and monitoring

Agents fail quietly — a confident wrong answer doesn't throw an error — so the bot has workflows around it whose only job is to watch it.

• Error handler (running). Every failed execution in the Hanami workflows is saved to an error log and raises an alert on Telegram.
• Per-turn telemetry. Each turn records the intent, latency, the tools used and whether it fell back — including a flag for mentioned a time without checking the calendar. In the newest version it's written before the DM is sent, so a failed send can't erase the record.
• Nightly evaluator and daily digest (built for the v3 cut-over, currently in staging). At 2 a.m. an LLM judge scores the day's conversations on tone, resolution, price hallucination, handoff and close; at 9 a.m. a digest with technical health, a per-agent breakdown and business numbers goes to Telegram.

Key decisions & trade-offs

The hard part

The agent offered appointment times that didn't exist.

How it was detected. The salon's owner reported it: the bot got the month wrong and offered slots that weren't actually free, then had to take them back.

Root cause. The earlier version handed the model raw calendar events and asked it to work out the free slots itself — subtracting busy events from opening hours, handling the time zone, and even keeping to a month that was hard-coded in the prompt. That's arithmetic, and a language model does arithmetic by guessing what looks right.

The fix. The model no longer computes anything about time:

1. A code step calculates the free 60-minute slots in the America/Mexico_City time zone: opening hours minus busy events, with a 24-hour minimum notice, over the next 14 days.
2. The list goes into the prompt with exact ISO timestamps and today's date, under one rule: you may only offer times that appear in this list.
3. To book, the agent copies the ISO value of the chosen slot into the booking tool — it never types a date on its own.

The calculation later became a reusable sub-workflow — the single source of truth for availability — and telemetry now flags any turn that mentions a time without having checked the calendar.

The lesson generalizes to any booking agent: the calendar is the source of truth, and the LLM is only the interface to it.

Results

Handling the salon's Instagram DMs since

Dec 2025

Book, reschedule and cancel, directly in the owner's calendar.

Major versions in production

3

The third made availability deterministic and added parameterized SQL and payload validation.

Website · Lighthouse SEO (mobile)

100 / 100

Lighthouse 13, 3 runs, September 2026.

Website · Lighthouse Best Practices (mobile)

100 / 100

Same runs.

Stack