LiveDesk AI: Claude
Chatbot for Live Houses

A customer in Osaka opens a live house’s website near midnight, wondering whether Saturday is free for the band they manage. The staff went home hours ago, so the question sits in an inbox until morning. By then the customer has asked another venue.

LiveDesk AI answers the moment the question arrives. It is a multi-tenant chatbot platform for Japanese live houses, the small live-music venues where the same inquiries come in again and again. A manager creates an account, types in venue details, adds FAQs, and connects Google Calendar. LiveDesk AI turns those three things into an assistant, powered by Claude, that replies in natural Japanese from behind one line of JavaScript on the venue’s own site.

Put simply, each venue trains its own assistant on its own information, FAQ wording, and live calendar, then pastes a single snippet to add a “Chat with us” widget. It answers questions about hours, tickets, access, and open dates around the clock, and never mixes one venue’s answers with another’s.

The Starting Point: One Snippet, Every Venue’s Own Answers

Live houses run on quick answers, but the people who give them are busy running shows. The brief was not one clever bot. It was a platform where any venue could stand up its own assistant without writing code.

• Answer the repetitive questions, day or night. Hours, prices, access, whether a date is free, often after closing when no staff are online.
• Answer schedule questions the customer cannot check. “Is this date available?” lives in a calendar only staff can see.
• Work for non-technical managers. Venue staff run events, not software, so anything that needs code sits unused.
• Learn each venue’s own wording. One venue says “available dates”, another “event dates” or “open slots”.
• Sound right in Japanese. A stiff or incorrect reply reflects on the venue.
• Install with one snippet. Managers cannot build chat widgets, so embedding had to be a single pasted line.

The Build: From Venue Details to a Live Assistant

A venue goes from nothing to a working assistant through a dashboard of forms and buttons. Getting a model to answer accurately in Japanese from one venue’s data, rather than guessing, is the substance of the Claude AI development behind the platform. Setup runs in order:

1. The manager signs up with email and password, or Google in one click.
2. They fill in Store Information, the venue’s details in its own words.
3. They add question-and-answer pairs in FAQ Training.
4. They connect Google Calendar with one OAuth click in Calendar Sync.
5. They try questions in AI Test before going live.
6. They click Generate Snippet and paste it in, and a “Chat with us” widget appears.
7. Once live, they review Chat History and add FAQs where answers fell short.

On the customer’s side it is just a chat. Someone opens the widget and asks in Japanese. The assistant pulls the venue’s own information and FAQs, reads the live calendar the way that venue has taught it to, and replies. The FAQ is where the real work happens: a manager can teach it that a day with no event means that day is available, or that “event dates” and “available dates” mean the same thing.

Technical Architecture

LiveDesk AI is multi-tenant: many venues, each with an assistant trained on its own data, all served by one system. The chat widget is the easy part. The engineering sits in grounding each assistant in the right mix of details, FAQs, and a live calendar, keeping tenants separate, and making setup trivial.

• AI answer engine. Claude, via the Anthropic API, writes the replies in natural Japanese, grounded in the retrieved venue facts and calendar.
• Application backend. Node.js with TypeScript runs multi-tenant auth, venue management, snippet generation, and chat orchestration.
• Retrieval. A pgvector store holds each venue’s embeddings, scoped to that tenant, so the assistant retrieves only that venue’s facts. Wiring a grounded assistant into a site is a common thread in AI chatbot development.
• Database. PostgreSQL stores venues, managers, FAQs, and chat logs with tenant isolation.
• Dashboard. A React front-end gives managers a simple, form-driven interface.
• Calendar and auth. The Google Calendar API supplies availability through OAuth; managers sign in with email or Google.
• State and cache. Redis handles sessions, rate limiting, and jobs.
• Chat widget. An embeddable JavaScript snippet renders the widget on the venue’s site.

The part worth engineering carefully was the join between the calendar and the assistant, because a raw calendar never says “available”. Events are pulled in real time, then read through the rules each manager wrote in the FAQ, so “is Saturday free?” comes out right for that venue. Because every venue’s data is scoped to its tenant, answers never leak between venues.

Challenges Solved

The hard problems were the ones that break a generic chatbot. Reading each venue’s calendar correctly was solved by teaching the rules in the FAQ rather than hard-coding them. Answering only from the right venue’s data was solved by scoping each tenant’s facts. Usability was solved by making everything a form or a button, so a venue goes live in minutes. Fluent Japanese was solved by grounding Claude in the venue’s own words and letting managers check replies in AI Test and Chat History. And when the assistant misses something, the manager adds an FAQ and the answer improves at once, with no developer involved.

The through-line matters for hospitality and venue businesses, where a question left unanswered after hours is often a booking lost to the next venue on the list.

The Result

The result is a platform where each live house has its own Japanese-speaking assistant that answers venue questions, reads the calendar the way the venue means it, and runs behind a single snippet, day or night. Managers who run more than one venue simply add another assistant, each trained and kept separate. This is the kind of work our Claude AI development services take on: an assistant grounded in a client’s own data, not a generic model.

The build works because it was designed around grounding and tenant isolation from the start. If you are planning something similar, you can hire agentic AI developers who have built this kind of grounded, multi-tenant assistant, or tell us about your venues and your calendar and we will map them to an assistant that answers in your customers’ language.

Frequently Asked Questions

What does LiveDesk AI do?

It gives each Japanese live house its own AI assistant that answers customer questions in natural Japanese, around the clock. Each venue trains it on its own information, FAQs, and Google Calendar, then embeds it on its website with one line of JavaScript.

How does the assistant answer schedule questions?

It reads the venue’s live Google Calendar, then interprets it using rules the manager wrote in the FAQ. Taught that a day with no event means that day is available, it answers “is Saturday free?” correctly for that venue rather than guessing.

How is one venue’s data kept separate from another’s?

LiveDesk AI is multi-tenant. Every manager, venue, FAQ set, calendar, and chat log is scoped to its own tenant, and each venue’s information is embedded in a vector store scoped to that venue, so answers never leak between venues.

Do managers need technical skills to set it up?

No. Everything is a form or a button: sign up, add details, connect Google Calendar with one OAuth click, and copy a snippet onto the site. There is no code to write, so a non-technical manager can go live in minutes.

What technology runs behind LiveDesk AI?

Claude answers through the Anthropic API. A Node.js and TypeScript backend runs the multi-tenant logic, a pgvector store grounds answers, PostgreSQL holds venues and chat logs, a React dashboard drives setup, the Google Calendar API supplies availability through OAuth, and Redis manages sessions and jobs.