Solution

AI Guide & Interpretation System Development

An AI guide and interpretation system extends or replaces on-site human interpretation with an AI guide, letting visitors ask questions in their own language and follow a tour path shaped by their own interests. It suits government venues and museums, corporate showrooms, exhibition and event reception, and tourism sites. The system delivers multilingual tours, visitor reception, and behavioral analytics, and can be deployed on cloud, private cloud, on-premise, or fully offline to meet each organization’s security requirements — keeping data inside your own boundary. Pumpkin VR builds on the Pumpkin AI Agent Platform, integrating venue knowledge bases, a content management back office, and on-site hardware such as handheld guides, touch displays, and visitors’ own phones — from PoC validation through full deployment and ongoing maintenance.

Venue types
4 types
Modules
AI Guide、AI Instructor、AI Concierge
Deployment
Cloud、Private Cloud、On-Premise、Fully Offline
Acceptance output
Acceptance clauses ready for contract
01

Where it applies

Interpretation needs differ sharply by venue. We tune knowledge-base depth, visitor flow, and hardware to each venue type.

  • Permanent and special exhibition interpretation
  • Multilingual reception for international visitors
  • Corporate showroom tours and briefings
  • Trade-show booth reception and live Q&A
  • Self-guided tours for tourism parks and attractions
  • Senior-friendly and accessible interpretation

Government Venues & Museums

An AI guide for permanent exhibitions, special exhibitions, and collection knowledge — filling volunteer-docent gaps while meeting the security and data-residency requirements typical of public procurement.

  • Collection and exhibit knowledge base
  • Docent workload relief and visitor distribution
  • Multilingual visitor reception
  • Fully offline option, data stays in-house

Corporate Showrooms

Hand brand story, product lines, and technical specifications to an AI guide — clients get answers on demand, and your sales team no longer has to narrate every visit.

  • Brand story and product-line interpretation
  • On-demand technical specification lookup
  • Client visit flow and briefings
  • Visitor interest and behavior logging

Exhibitions & Event Reception

A reception agent serves multiple booth visitors at once — answering repetitive questions and collecting contact details so your team can focus on real conversations.

  • Concurrent Q&A for multiple visitors
  • Automated answers to repeat questions
  • Lead and requirement capture
  • Short-term rental for show periods

Tourism Sites

Self-guided tours for parks, attractions, and heritage streets — visitors scan a code with their own phone and start, with no device to check out.

  • Scan-to-start self-guided tours
  • Point-of-interest and route narration
  • Free language switching
  • Lower off-peak staffing cost
02

Three Core Modules

The system is assembled from three modules of the Pumpkin AI Agent Platform — adopt one, or combine them to fit your venue.

AI Guide

The visitor-facing interpreter. It answers questions about exhibits and spaces, strings together tour routes, and unfolds deeper background knowledge as visitors dig in.

Explore the platform →

AI Instructor

The staff-facing trainer. Used for onboarding docents and new interpreters, rehearsing standard interpretation content, and answering questions about venue operating procedures.

Explore the platform →

AI Concierge

The front-desk and booth agent. It handles visitor enquiries, wayfinding, session and facility questions, and group reservations.

Explore the platform →
Technical capabilities
  • High-fidelity development with Unreal Engine
  • Cross-platform development with Unity
  • Full VR / AR / MR device support
  • AI-driven interaction and learning analytics
  • Motion simulators and custom hardware integration
03

Deployment Options

The same system deploys wherever your security policy and network conditions require — you do not trade features for compliance.

Cloud

A cloud environment we operate. Fastest to launch — best for short-run exhibitions and the PoC stage.

Private Cloud

Deployed inside your own cloud tenancy, so data stays within the governance boundary you already operate.

On-Premise

Deployed in your own server room, with your IT team holding the hosts and access control.

Fully Offline

An on-site host that runs with no external network connection at all — for closed network segments and venues under strict egress restrictions.

Data residency: with private cloud, on-premise, or offline deployment, all interpretation content, knowledge bases, and visitor interaction records remain inside Taiwan in the environment you designate — nothing is sent to overseas servers. The exact deployment location, retention period, and access controls are confirmed in writing before build.

04

Language & Device Support

Multilingual Tours

One knowledge base outputs multiple languages — update content once and every language follows, with no per-language scripts to maintain.

  • Traditional Chinese, English, Japanese, and Korean are the common set
  • Additional languages on request
  • Voice input and spoken playback
  • Locked translations for proper nouns and house terminology

Device Support

Runs on everything from a visitor’s own phone to the touch hardware you already own — no hardware refresh required to adopt it.

  • Visitor’s phone browser (scan to start)
  • Touch displays and standing kiosks
  • Handheld guide devices and tablets
  • AR / VR headsets
05

Where this engine already runs

A guide is one use of this conversational engine. Rather than describe it only in specifications, here is where it currently runs, so you can judge for yourself whether it holds up in a real environment.

Swipe to see all columns

WhereStatusEngine capability evidenced
Kaohsiung Veterans General HospitalDeployedHEKA running inside a medical center: long, multi-turn, user-led conversational interaction.
Sports-skill MR training systemDeliveredA real-time AI coach with customizable persona and voice, answering strictly inside a defined body of knowledge rather than improvising.
Cheng Shiu University AI adoption programmeIn progressMethodology for adoption and behavior change at institution scale: assess, seed, diffuse, then hand operation back to the organization.
MR capture. The coach is a 3D character we built, standing in the user’s actual room — the background is unretouched, which is precisely what separates passthrough MR from a pre-built virtual set. Swap the knowledge base for a venue’s and the same engine becomes a guide.

⚠️ The table above evidences engine capability, not delivered venue tours — we have no public guide deployment yet. What actually decides a guide project is the completeness of the knowledge base and how acceptance is defined.

06ACCEPTANCE PROTOCOL

How acceptance criteria are set

Almost every acceptance dispute over a guide system comes from never having defined what "answering well" means. We agree the test set and pass thresholds with you at the contract stage, and write each one as a clause that can be measured and re-tested. You set the threshold figures based on your venue and visitor mix; we supply the measurement method and the form of evidence.

Swipe to see all four columns

ItemTest methodPass conditionEvidence
Answer accuracyScored question by question against a test set you supply and approve, judged by your own staffPass rate set by you and written into the contract; re-tested on site at acceptancePer-question score sheet and a list of incorrect answers
Offline readinessAll guide functions tested with the external network disconnected, with response times loggedNo function may be missing while disconnected; the response-time ceiling is set by you from your actual visitor flowDisconnected-test record and a response-time measurement sheet
Multilingual content qualityEach language proofread by you or your designee, with proper nouns checked against a pinned glossaryProofreading pass rate set by you; no proper noun may render inconsistently across languagesPer-language proofreading record and the pinned glossary
System availabilityOutage count and duration during opening hours, logged across an agreed observation periodObservation-period length and the availability threshold are stated in the contractSystem log for the observation period with an explanation of each outage
HandoverYour staff perform a real content update in the admin console; we may not do it for themThe handover checklist is signed item by item; unsigned items do not count as completeSigned handover checklist and training record

These fields can be lifted directly into a requirement document or acceptance clause. The threshold figures are left for you to fill in because a standard that fits a permanent exhibition may not fit a special exhibition or an outdoor site — a number we invented would serve neither side.

07

Representative cases

View all work →
08

Frequently asked questions

How is the cost of an AI guide system structured?

Cost has four parts: knowledge-base construction (turning your existing scripts, catalogues, and exhibit data into a usable knowledge base); system build and customization (interface, visitor flow, house style, and integration with existing systems); the deployment environment (cloud, private cloud, on-premise, or offline — each priced differently); and post-launch content updates and maintenance. The figure depends on venue scale, number of languages, and module mix — we provide an itemized quote after a discovery session.

How long does deployment take?

A single-venue PoC is typically usable on site in 4–6 weeks. A full production build — including knowledge-base preparation, multilingual proofreading, and hardware integration — generally runs 3–5 months. The largest variable is the state of your interpretation content: complete, already-digitized scripts shorten it noticeably, while a fresh inventory of exhibit data must be planned into the schedule.

Which languages are supported, and does adding one mean rebuilding?

The common set is Traditional Chinese, English, Japanese, and Korean, extendable to your actual visitor mix. Because all languages are generated from one knowledge base, adding a language requires no content-architecture rework and no separate per-language scripts. Venue names, exhibit titles, and proper nouns can be pinned to fixed translations so the same term never renders inconsistently across languages.

Can it run in areas without network access?

Yes. In offline deployment the language model and knowledge base both live on an on-site host, so the exhibition area needs only a local network and no external connection at all. This suits agencies on closed network segments, underground galleries, and outdoor sites with unreliable signal. Content updates in offline environments are applied as scheduled batch syncs by your own staff from the admin console.

Will our data be sent overseas?

Not with private cloud, on-premise, or offline deployment. Interpretation content, knowledge bases, and visitor interaction records all remain inside Taiwan in the environment you designate, with nothing sent to overseas servers. Before the build we confirm the actual deployment location, retention period, and access controls in writing, in a form suitable for security review and procurement documentation.

How far does customization go?

Interface visuals, visitor flow, interpretation tone and persona, knowledge-base structure, and integration with your existing ticketing, CMS, booking, and membership systems are all customizable. The persona can reuse an existing venue mascot or be designed fresh. What cannot be customized is the underlying capability of the language model itself — but we constrain it through the knowledge base and answering rules so it answers only from the material you supply, rather than saying something unverified on the exhibition floor.

How is content updated after launch, and who maintains it?

Your staff update exhibit descriptions, special-exhibition content, and FAQs themselves through the admin console — no engineering involvement and no redeployment required. For maintenance there are two paths: an annual maintenance contract with us (covering system updates, model tuning, and troubleshooting), or admin-console training at handover so your own IT team operates it, with a technical support channel kept open on our side.

What acceptance criteria can we set?

Commonly used measurable criteria include: correct-answer rate on questions covered by the knowledge base (scored against a test set you supply), average response time, proofreading pass rate for each language, system availability, and visitor usage plus tour-completion rate. We recommend agreeing the test set and pass thresholds with you at the contract stage, so acceptance does not turn into a disagreement about what counts as "answering well."

Need material you can put into a spec?

Tell us the venue type, visitor mix, and your internal approval process. We will provide an architecture description, a draft of acceptance clauses ready for contract, and a deployment recommendation that meets your security requirements.