Where does the visitor fit in all this?

They're the protagonist. The strongest immersive work decenters the narrative and blurs the line between the person and the environment. Active participation is what gets you to deep immersion.

I think that matters more now than ever. Our phones already give us endless private screens. What a physical installation can do that a phone can't is put people in shared contact with technology: in the same room, reacting to the same thing, together.

Technology is obviously central. How do you think about its role, especially with AI?

On every project I've worked on, something had to be built, and what "building" means keeps changing. Early on it was getting a broadcast signal off a remote site. Then it was architecture that responds to people in real time. More recently it's been real-time animation pipelines feeding generative AI systems.

AI is the latest shift, and it's a big one. As it removes the cost of producing pixels, the value moves to whoever supplies the intent and context. It's a bit like what the printing press did to the scribe. The craft doesn't disappear. It moves upstream, toward deciding what's worth making and why. That's the part of the work I care about most.

You've started describing ND's work as precursor for building effective AI agents. What's the connection?

When people talk about an AI agent, they usually mean the model. But the model is only one piece. Around it sits a harness: the layer that takes in what people say or do, routes it to the model, checks what comes back, and delivers it to a screen, a speaker or another system. That harness is where most of the real engineering lives, and it looks a lot like what we've been building in physical spaces for years.

An interactive installation is a system that manages state, inputs and latency in a live, unpredictable environment, full of people who won't do what you expect. An agent harness is the same kind of system. The materials are different, but the discipline is very similar.

Walk us through how that maps.

Start with the guest journey. In experience design you map how someone moves from one room to the next, what they're likely to try, and how the space should respond. In a harness you're mapping agent states and routing user intent instead of rooms, but it's the same kind of thinking. Designing an interface that keeps a member of the public from breaking an installation isn't far from designing the guardrails and system prompts that keep a user from pushing a model off the rails.

Then there's systems integration. Much of my career has been building signal backbones: getting sensors, cameras, LED walls and audio onto one network with latency low enough that the room feels alive. A harness is a backbone too. Voice, text and API calls come in, pass through the model, get checked, and go back out to the person. The concerns are the same (latency, throughput, network security), just with different payloads. We've also built systems that run entirely on a client's own hardware, sometimes air-gapped, to keep their data in-house. That's increasingly what organizations want from private AI deployments.

What about risk?

That's the part I think transfers most directly. Live delivery is unforgiving: it happens once, in public, with no patch. If tracking drops in the middle of a launch event, everyone sees it. A public-facing AI agent has the same exposure. People will try to break it, inject prompts, or bait it into saying something it shouldn't.

So we bring the same defensive habits we use for a live show. We validate inputs, constrain what the model is allowed to do, and always build a deterministic fallback, so a problem degrades gracefully instead of failing in front of an audience. No system is unbreakable, but you can design one where failure is quiet and recoverable.

Is this theoretical, or are you already doing it?

We're already doing it. For an AI-driven vehicle launch, we combined speech recognition, an LLM-based intent classifier and a live generative graphics pipeline into one experience that had to perform in real time, in front of guests. It was a live show and an agent harness at the same time.

From the room to the agent

Guest journey → intent mapping and state control

Sensors and presence tracking → multimodal inputs (voice, image)

Signal backbone and switching → orchestration and tool-calling pipelines

Local execution and firewalls → on-premises or sandboxed model deployment

Show-day failover → guardrails and deterministic fallbacks

What separates installations that work from ones that fall flat?

A clear design vision, the right mix of people around the table, and a willingness to adapt the idea to the context instead of forcing a signature move into every room. Cutting-edge technology helps, but only when it's in service of the vision. Plenty of expensive installations fail because the tech was the idea.

How important is testing and iteration?

It's essential, and it needs to start earlier than people think. One thing I watch closely now is that AI renders can look finished long before anyone knows whether the thing is buildable. In early phases I'd rather use deliberately rough, low-fi visualization and get it in front of real people while it's still cheap to change.

Context matters here too. On a medical innovation lab we delivered, the audience wasn't one-time visitors. It was staff and invited guests who'd be in that space every day. An effect that delights someone once can grate by the tenth day. You only learn that by testing with the people who will actually live with it.

Does the approach change between a brand activation, a museum and a corporate event?

The fundamentals don't, but the goals do. Corporate events, entertainment, retail, sports and cultural institutions each have a different audience and a different definition of success. I start with who's in the room and what the organization needs from it, then design backward from there.

Last question: what kind of team does this take?

It rarely fits inside one discipline. You need interactive engineering, content and animation, UX, hardware and systems, and experience design, and every project weights them differently.

That's how Neon Diorama is built. We're a generalist studio with a precision workflow: we identify the building blocks a project actually needs, then bring in the right people to fill the gaps. Often that means working alongside a creative studio that wants to stay focused on the creative work, while we handle the production, systems and technical detail that would otherwise pull them away from it.