How an AI-powered Star Wars character led to the birth of Intari
In September 2022, I was fulfilling a lifelong dream.
I was inside the Star Wars universe.
I’ve written about this experience before, but Disney’s Star Wars: Galactic Starcruiser was an immersive attraction far ahead of its time—and far beyond what any marketing campaign could have adequately promised.
Guests aboard the Halcyon, the fictional galactic cruise ship where the story took place, could interact with a fully realized animatronic alien, a physical droid roaming the ship, and an animated AI-powered character named D3-09 who appeared on a screen inside each passenger cabin.
Within the story, D3 was a logistics droid: a combination of ship’s steward, historian, concierge, and confidante. She could answer questions, tell stories, inform passengers about events unfolding around the ship, and even conduct an impromptu vow-renewal ceremony for guests celebrating an anniversary.
Put succinctly, she was incredible.
But she also had obvious limitations.
During a long conversation, you might use a word she didn’t understand or ask a question she wasn’t prepared to answer. When that happened, D3 would be forced to fall back on some variation of “I’m sorry, but I didn’t understand you.”
Common slang could confuse her. Sarcasm could confuse her. Questions that ventured too far outside her anticipated subjects could bring the conversation to an abrupt dead end.
The longer you spoke with her—and especially if you tried to test the boundaries of what she could do—the easier it became to find those dead ends.
Fortunately for both D3 and the guests, Disney’s writers had cleverly incorporated many of her limitations into her personality. D3 was a logistics droid aboard a civilian cruise liner that suddenly found itself at the center of a battle between good and evil. When she appeared confused, overwhelmed, or ill-equipped to handle the chaos unfolding around her, that reaction made sense within the story.
Her technical limitations became character traits.
That was brilliant storytelling. It protected the illusion even when the technology could no longer sustain it.
But for people like my wife Candice and me, with our overlapping interests in narrative structure, character development, design, and technology, D3 also raised a much larger question.
Could we make a character like this feel even more real?
What was actually happening behind D3-09?
We didn’t have access to Disney’s technology, and we still don’t know the complete architecture behind D3-09.
She may have used carefully authored dialogue trees, intent recognition, retrieval systems, machine learning, or some proprietary combination of technologies developed by Disney Imagineering. Our experience with her was not a technical audit. It was simply an interaction between two passengers and a remarkably ambitious fictional character.
So it would be unfair to declare that D3 was built using any one particular methodology.
What we can say is that our conversations with her exposed a familiar limitation: she appeared capable of operating within a prepared body of knowledge and behavior, but she struggled when we moved beyond it.
We walked away from the Galactic Starcruiser asking a deceptively simple question:
How could we build a character that didn’t reach the edge of itself so quickly?
A little over a year later, Candice and I were enrolled in postgraduate studies in Artificial Intelligence and Machine Learning at the UT McCombs School of Business here in Austin, Texas.
In our classes, we studied contemporary methods for building agents and chatbots. One of the most common was a methodology called RAG.
What RAG actually does
RAG stands for retrieval-augmented generation.
In simplified terms, RAG allows a generative AI system to retrieve relevant information from an external source and use that information as context when producing a response.
Imagine that we create an AI cooking assistant and provide it with a collection of recipes. If someone asks, “How do I bake a carrot cake?” the system can search that collection, retrieve the most relevant passages, and give them to the generative model. The model then uses that material to formulate an answer.
This is extraordinarily useful.
RAG can give an AI access to private company information, current product documentation, technical manuals, educational materials, fictional histories, medical references, or any other specialized information that was not reliably included in the model’s original training.
It can help ground the model’s responses in approved material. It can make information easier to update. It can reduce unsupported invention and give users access to the sources behind an answer.
But RAG is not simply a library of scripted responses, and a generative model does not necessarily stop functioning when retrieval fails. Depending on its instructions and guardrails, it may answer from its general knowledge, explain that the available sources do not contain the answer, ask a clarifying question, or—if the system is poorly governed—make something up.
That distinction matters.
Because the more we studied RAG, the more we realized that retrieval was not actually the problem we wanted to solve.
RAG could give a character knowledge.
Knowledge alone could not make something a character.
Knowledge is not character
Suppose we wanted to build a new logistics droid.
We could provide it with every available piece of information about the Halcyon. We could give it the ship’s complete history, technical specifications, passenger itinerary, crew manifest, and thousands of pages of Star Wars lore.
That system might know exactly where the engineering room was located. It might know what was being served for dinner. It might even know the complete history of the ship’s previous captains.
But none of that would tell the system how it felt about a particular passenger.
It would not determine whether the droid trusted that passenger more after their third conversation than it had after their first. It would not give the droid something it wanted. It would not make the droid nervous about the First Order, protective of a crew member, suspicious of a smuggler, or reluctant to reveal a dangerous secret. It would not give the character a reason to speak.
RAG could become the character’s library. It could not become the character’s inner life. That was the idea we eventually killed: not retrieval itself, but the belief that retrieval was enough.
Problem one: identity
The first problem was personality—but even “personality” felt like too small a word.
A character is not just a collection of verbal habits. Giving a chatbot a system prompt that says “be friendly, charming, and slightly sarcastic” may affect the style of its language, but style is not identity.
A convincing character needs rules and boundaries, but it also needs motivations.
What does this character want? What is it afraid of? Whom does it trust? What information is it trying to obtain or protect? What would cause it to change its mind? How do its goals align—or conflict—with the goals of the person speaking to it?
Consider Sherlock Holmes and Doctor Watson. Both characters might possess information about the same mystery, but they would not discuss it in the same way. They would notice different details, express different concerns, and pursue different objectives.
Their knowledge would influence the conversation, but their motivations would shape it.
The same principle applies to a business assistant.
Does the assistant want to resolve a customer’s problem? Help the customer make a decision? Teach the customer how to perform a task? Recommend a product only when it is genuinely relevant?
Those motivations become visible through behavior. They determine whether an interaction feels helpful, manipulative, compassionate, frustrating, or indifferent.
Our system would need to remain governed without being trapped inside scripted responses. We needed to define who and what a character was, give it a purpose, establish firm boundaries, and still allow it enough freedom to improvise.
Problem two: continuity
The second problem was memory.
Human conversations work because each statement exists in the context of everything that came before it.
In conversational AI, we often refer to each exchange between a human and a model as a “turn.” At every turn, the AI needs enough information from previous turns to understand what is happening now.
But simply retaining a transcript is not the same as remembering.
A useful memory system must determine what matters.
Was a particular statement a passing remark, or did the user reveal something important about themselves? Was a promise made? Did the relationship change? Did the character learn something that should affect a future decision? Has the character already told this story?
The system also needs to remember its own behavior. Otherwise, a character may repeat itself, contradict itself, disclose the same secret twice, or behave as if an important moment never happened.
For a character to feel persistent, memory cannot merely be stored. It must influence what happens next.
Problem three: performance
Human communication contains far more than words.
We interpret pacing, hesitation, posture, expression, tone of voice, and silence. We recognize the difference between someone saying “I’m fine” cheerfully and saying those same words while avoiding eye contact and struggling not to cry.
If our characters might eventually power audio performances, animated avatars, games, or real-time rendered experiences, dialogue alone would not be enough.
The system needed to supply what we began calling stage direction: structured indications of what the character was doing beyond the spoken words.
Is the character avoiding eye contact? Is it sighing heavily? Is it struggling to conceal excitement? Does it pause before answering? Does it appear frightened, impatient, amused, or haunted?
When I use words such as “feelings” here, I am not claiming that these systems are conscious or experience emotion in the human sense. I’m describing modeled internal state: information that represents a character’s current disposition and changes how that character behaves.
That state needed to persist. A character could not appear devastated in one sentence and inexplicably cheerful in the next unless something had happened to cause the change.
Then there was the challenge of group interaction.
What if a human was speaking with several characters at once? Each character would need its own identity, knowledge, memories, motivations, emotional state, and nonverbal behavior.
Those characters might agree. They might interrupt one another. They might compete for the user’s trust. They might withhold information from one another. Their goals might conflict with the goals of the human—and with each other.
At that point, we were no longer talking about a chatbot.
We were talking about a cast.
The birth of Intari
Those questions led Candice and me to begin developing Intari.
Intari is our approach to building persistent, intuitive, and improvisational AI characters. Rather than treating personality as a few paragraphs placed inside a system prompt, we treat a character as a changing combination of identity, goals, rules, relationships, memories, emotional state, knowledge, and performance.
Retrieval can still be part of that architecture. If a character needs access to specialized or current information, RAG remains a useful tool.
But it is only one tool.
The character should not cease to exist when retrieval has nothing useful to provide. It should be able to ask questions, reason within its boundaries, respond according to its motivations, and remain recognizably itself throughout the interaction.
That became the dividing line for us.
We were no longer trying to build a bot that could answer in character.
We were trying to build a character capable of having an experience with someone.
Once we started approaching the problem that way, the idea began to snowball.
Try it out yourself:
Below you’ll find an example Intari chat using characters from Candice’s novel, “Dao of Zed: Dead Weight.” At this point, near the beginning of the book, the protagonist, Andy, has yet to take his “call to adventure” and begin the quest that will lead to the apocalyptic fall of society the book centers around. So the context for these two Intari is very specific: we find the overweight hero and his thinner best friend (Brent) hanging out in their shared college apartment. Brent and Andy have different opposing motivations (Andy wants to leave for a government-run medical trial for a miracle weight loss treatment and Brent wants him to stay home and lose weight in more “traditional” ways), and they will try to convince you, the audience, to side with them in very different ways.
Nova and Amplify ADHD
Nova is an Intari being developed as part of an application called Amplify ADHD.
She is designed to help people with ADHD plan their day, prioritize tasks, and work through common executive-function barriers. Her role is not merely to retrieve productivity advice. She needs to respond to the individual, understand the immediate context of the conversation, and adapt her approach when a strategy is not working.
Someone struggling with task initiation may need a different kind of support than someone who is overwhelmed by too many priorities. Someone having a difficult emotional day may need compassion before they need another checklist.
Nova is not intended to replace a qualified mental-health professional or provide clinical diagnosis. Her purpose is to offer practical, supportive coaching within appropriate boundaries and direct users toward professional or crisis resources when their needs extend beyond those boundaries.
This application asks whether an AI character can provide assistance that feels responsive without pretending to be human.
Avery and Stay Positive
For the financial application Stay Positive, Candice created Avery, an Intari financial coach.
Avery is designed to help individuals and families understand their options for getting out of debt, paying bills on time, establishing healthier financial habits, and planning for long-term goals such as a child’s education.
Financial information can be intimidating, and shame often prevents people from seeking help. The challenge is not only to present accurate information, but to do so through a character capable of being encouraging without becoming dishonest, judgmental, or unrealistically optimistic.
Avery explores how a governed character can maintain a supportive identity while helping users face difficult practical realities.
Continuum
Continuum began as what we jokingly described as a “marketing organization in a box.”
The proof of concept contains multiple conversational rooms, each populated by a different team of Intari characters.
In the first room, the user presents a business idea to a C-suite of AI characters representing different executive perspectives. A financial leader may question the economics of the proposal. A technical leader may identify operational challenges. Other characters may focus on risk, customers, or organizational readiness.
The user can then carry the important outcomes of that discussion into another room, where an executive marketing team turns the proposal into a cohesive strategy.
From there, the strategy can move into a final room populated by creative professionals and project managers who plan its execution.
At the end of the process, the user can download a document that traces the idea from its original pitch through strategic refinement and campaign planning.
Continuum is not simply several chatbots placed beside one another. It explores how multiple characters with different responsibilities and motivations can collaborate, disagree, and build upon one another’s work.
Intari Realms
Intari Realms extends the idea into interactive entertainment.
It is a game platform where players can enter a world populated by 20 or more Intari characters. Each character has an identity, motivations, information, relationships, and a role within the larger story.
Our first Intari Realms experience was a murder mystery.
Players questioned a collection of AI-powered suspects to determine who committed the crime, how it was done, and why. Asking productive questions could persuade a character to reveal a clue, but the characters did not all have the same knowledge or reasons for sharing it.
Discovered clues were stored in the player’s inventory until the player was ready to make a formal accusation.
This experience tested many of our original questions simultaneously: persistent character identity, conflicting motivations, conversational memory, improvisation, world state, and the ability to keep many AI characters operating inside one coherent narrative.
In other words, it brought us surprisingly close to the question D3-09 first inspired.
What would it take to build a character that could live inside a story without being limited to a predetermined conversation?
Intari Mark
On this website, visitors can also speak with an Intari version of me.
Intari Mark is my interpretation of a sales chatbot, although his goal is not simply to push a product. His purpose is to help visitors understand my background, experience, projects, and perspective through a conversational interface that resembles the way I actually communicate.
People who know me personally have described the experience as eerily similar to speaking with the “real” me.
That does not mean Intari Mark is me, of course. He is not conscious, and he does not possess my complete knowledge or judgment.
But his consistency demonstrates what can happen when an AI experience is designed around identity and motivation rather than a collection of frequently asked questions.
So, is RAG really dead?
No.
Not literally.
RAG remains an effective way to give AI systems access to reliable, specialized, private, or current information. It can ground an answer in approved sources and help reduce unsupported claims. Intari characters can use retrieval just as other knowledge-intensive systems can.
What we no longer believe is that retrieval alone can create a convincing character.
A character needs knowledge, but it also needs identity, memory, motivation, relationships, emotional state, and the ability to behave within a changing world.
It needs to do more than find the right answer.
It needs to have a reason for giving that answer—and sometimes the answer should change because the character has changed.
D3-09 did not fail us. Quite the opposite.
She showed us how close technology could come to creating the experience of meeting someone who had previously existed only in fiction. Then she showed us the edge of that experience and left us wondering what might exist beyond it.
That question became Intari.
It became Nova and Avery. It became Continuum, Intari Realms, and Intari Mark. Each experiment revealed another part of the same larger problem: how to build AI interactions that feel less like navigating a menu and more like meeting a persistent presence.
So yes, RAG is dead—but only as the soul of the machine.
And perhaps what Candice and I really killed was the idea that knowing things is the same as being someone.