A Conversational AI Museum Exhibit That Stays PG
Impekable designed and built an AI voice avatar exhibit for a major natural history museum and science institution in San Francisco: three animated animal characters that visitors can walk up to and talk with in real time. The project covered the full build, from character voices to a content safety system, a fact management dashboard for non-technical staff, and an automated testing pipeline that ran more than 3,000 test conversations before launch. The exhibit is deployed and pending live launch.
The Brief
A world-renowned natural history and science museum wanted to bring one of its most beloved exhibits to life. Not through a screen, not with a pre-recorded audio loop, but with a real, conversational AI character that visitors could walk up to and actually talk to.
The concept: three animated animal characters, each housed in a kiosk that looked like an arcade machine, each capable of holding a two-way conversation about their lives, habitats, and the science behind who they are.
The audience: families, kids, school groups, and the curious public. Which meant every answer had to be accurate, every response age-appropriate, and nothing could go sideways in a crowded museum hall.
Simple concept. Enormously complex build.
The Problem
Most AI voice demos are easy. You spin up a voice model, write a system prompt, and in four or five hours you have something that feels magical. The hard part isn't the demo. The hard part is making it work in the real world: for thousands of strangers, on hardware you don't control, in an environment full of background noise, unpredictable questions, and zero tolerance for embarrassment.
This project had all of that, and more. The museum needed a conversational AI exhibit that could:
Answer factual questions accurately, drawing only from vetted, approved sources
Stay completely PG, even when visitors tried to get it off track
Sync audio output with animated character lip movements in real time
Run on modest, non-GPU kiosk hardware in a public space
Be managed and updated by a non-technical museum staff team
None of those requirements is hard in isolation. Together, they required a system built from scratch, tested relentlessly, and designed to handle the 5% of conversations no one planned for.
Before We Started
Before this project, the exhibit used static, pre-recorded audio:
No conversation. Visitors pressed a button and heard a fixed clip.
No follow-up. There was no way to ask what a sloth actually eats or why it moves so slowly.
No engagement. The experience was passive, and the museum knew it.
The museum wanted a museum AI experience that could hold a real exchange, answer genuine curiosity, and stay in character no matter what a seven-year-old decided to ask. That is a completely different product from a looping audio file. Everything in this case study was built from that starting point.
Why This Was Hard
You're Telling an LLM Not to Be an LLM
The single biggest technical challenge in AI voice for public-facing exhibits isn't the voice. It's the knowledge constraints. Large language models are trained to give answers, and they will confidently reach for something plausible even when they don't know the actual answer. In a museum, that isn't acceptable.
Impekable spent significant time building a fact system that served as the sole source of truth for each character. Every fact an avatar could reference had to be pre-approved, formatted for quick retrieval, and tested to confirm it surfaced the right answer under varied question phrasing. Getting an LLM to say "I don't know" convincingly, without breaking character, is harder than it sounds.
Hardware Reality vs. Demo Reality
The museum's kiosks were mini PCs running Windows with no graphics card. The particle animation effects the museum had requested, which it considered central to the exhibit's aesthetic, consumed nearly 100% of CPU on a brand-new MacBook Pro. On the kiosk hardware, that was a real problem.
Impekable laid out the tradeoff clearly: keep the particles and risk performance failures, or scale them back and keep the exhibit running. The museum chose to keep them. The team built around that choice as well as it could while being honest about the risk.
Safety at Scale in a Public Space
One inappropriate response in a kids' science museum isn't just a bug. It's a headline.
The team built a two-tier content safety system. Soft redirections handled questions that brushed against sensitive topics, letting the character gently steer the conversation without hard-stopping. Hard stops were reserved for anything crossing into self-harm territory or deliberately adversarial inputs. Both tiers were pre-scripted and tested, so responses fired instantly without generating new text on the fly.
In practice, 95% of questions were predictable: what do you eat, where do you live, why are you slow, how do you sleep? The system was built for the other 5%.
What We Built
Three Fully Voiced AI Characters
Slomo the Sloth, Mel the Whale, and Daisy the Bee, each with their own voice, personality, and knowledge base. Each character was built with ElevenLabs AI Voice Agents, and all three were independently tested and deployable.
A Custom Fact Management Dashboard
Because the museum team was non-technical, Impekable built a web-based admin dashboard that gave staff direct control over the characters' knowledge. It works like a simple approval queue: staff review AI-generated fact drafts, approve or reject them, and push approved facts to ElevenLabs with a single action. No code required.
The dashboard also includes conversation telemetry, sentiment scoring, safety flag monitoring, and the ability to replay and review individual sessions. Every conversation is automatically pulled in and logged as it ends.
An Automated AI Testing Pipeline
This was one of the most technically significant pieces of the project, even though no visitor ever sees it.
Before any system prompt change could go to production, it had to pass an automated testing loop. AI agents acted as simulated visitors, generating varied, realistic conversation scenarios. Other AI agents acted as judges, evaluating every response against the approved knowledge base and a defined set of quality criteria.
More than 3,000 test conversations were run this way over the course of the project. Every system prompt change triggered a new test run, because in voice AI a two-word change in one instruction can create unexpected behavior elsewhere. The pipeline caught those regressions before they reached a real visitor.
Pre-Scripted Response Architecture
Response speed matters in voice. A delay of even 200 milliseconds before the first word creates an uncanny pause that breaks the illusion of conversation. To eliminate it, Impekable pre-generated 50 variations of each opening greeting, vetted them, and served them from a pre-loaded cache. The same approach covered goodbyes and common redirection phrases, giving near-instant response onset for the most frequent interaction patterns.
Full Handoff Documentation
Impekable prepared comprehensive handoff documentation so the museum team could confidently manage the exhibit after delivery: a clear walkthrough, a GitHub transfer, and a setup guide written for a non-technical team. We also walked the team through the handoff and provided supporting materials to keep the transition smooth.
Meet the Characters
Results
The exhibit is deployed and ready for a multi-month public run, on infrastructure that costs less than $10 a month.
3 production-ready AI avatar characters, each with a distinct voice and personality
3,000+ automated test conversations run before deployment
A fact management system giving non-technical staff full editorial control over character knowledge
Built-in safety architecture covering both soft redirections and hard content stops
Full telemetry and session review capability
Total infrastructure cost under $10/month
Exhibit scheduled for a 3 to 4 month live run, with an estimated hundreds of visitor interactions per day
Project ready to go live pending hardware stabilization at the museum
"I think it's been a great success though! The people using it don't care that it's AI, they are loving the experience and we continue to keep building on it as a direction”
What We Learned
AI demos are fast. Production is slow. The first working prototype came together quickly. Getting the system to handle any visitor, any question, and any hardware condition, without fail, took five weeks of careful iteration. That isn't a problem. It's the actual product.
Test infrastructure is a project deliverable. The automated testing pipeline wasn't originally in scope. It became one of the most valuable things built. Any AI voice product going into a high-stakes environment needs systematic testing, not just manual QA.
The client dashboard is as important as the product. If a non-technical team can't manage the system, the system goes stale. The fact approval workflow lets the museum keep its characters accurate and current without ever filing a support ticket.
Telling an LLM what not to say is a real discipline. Guardrails in public-facing AI aren't a checkbox. They're a craft. Every restriction has to be tested across dozens of phrasing variants to confirm it actually holds. Impekable now has a repeatable process for this.
What's Next
This exhibit represents something genuinely new. It's believed to be one of the first open-air ElevenLabs deployments at this scale, meaning the character audio plays in a shared physical space rather than through a private headset or phone line.
The tooling built for this project (the fact management system, the testing pipeline, and the response caching architecture) is reusable. Future AI avatar exhibits for zoos, aquariums, science centers, or cultural institutions can be built faster and more reliably because of this groundwork.
Glossary
AI voice avatar: An animated character that speaks with a synthesized voice and holds real-time, two-way spoken conversations, powered by a large language model.
Large language model (LLM): An AI model trained on large volumes of text to understand and generate language. LLMs power the conversational responses of AI voice agents.
System prompt: The standing set of instructions that defines an AI character's personality, knowledge boundaries, and behavior rules for every conversation.
Guardrails: Rules and controls that limit what an AI system can say or do, used to keep responses accurate, safe, and on-topic.
Knowledge base: The approved set of facts an AI character is allowed to draw from when answering questions. In this project, it served as the sole source of truth for each character.
Soft redirection: A pre-scripted response that gently steers a conversation away from a sensitive topic without ending it.
Hard stop: A pre-scripted response that immediately blocks a conversation path when input crosses a defined safety line, such as self-harm topics or deliberately adversarial prompts.
AI judge: An AI agent that evaluates another AI's responses against defined criteria. Here, AI judges scored test conversations against the approved knowledge base.
Response caching: Pre-generating and storing common responses so they can be delivered instantly instead of generated in real time.
Open-air deployment: An AI voice deployment where audio plays aloud in a shared physical space rather than through a phone line or private headset.
Frequently Asked Questions
How do museums build an AI voice avatar exhibit?
What does an AI museum exhibit cost to run?
Does a museum need a technical team to run an AI voice exhibit?
How do you keep an AI exhibit safe and age-appropriate for kids?
Is an AI voice exhibit a good fit for zoos, aquariums, and science centers?
The Takeaway
A museum exhibit on modest hardware, in a room full of unpredictable visitors, with no technical staff on site, is one of the hardest possible environments to deploy conversational AI. It forces you to solve every problem you'd otherwise defer. You can't push a hotfix while a school group is standing in front of the kiosk.
That's the real lesson for AI exhibit design. Impekable built something that works under those conditions, not because the tech is simple, but because the team treated every constraint as a real problem worth solving properly.
Impekable is a UX and AI product studio. We design and build digital products that work in the real world, not just in demos. If you're thinking about AI voice for a public-facing experience, we'd like to talk.
