What happens when an old fortune-telling app becomes a place for private reflection?
For more than fifteen years, people knew my work through an unlikely piece of software: a Wong Tai Sin fortune-stick app.
It began with a simple problem.
People visiting Wong Tai Sin Temple would draw a fortune stick, receive a number, and then try to understand what it meant. A mobile application could make that process easier. Enter the stick number. Read the interpretation. Carry the reference in your pocket.
That was enough in 2009.
It is no longer enough in 2026.
The arrival of generative AI created an opportunity to rethink what this kind of application could actually do.
The obvious idea was to add a chatbot.
That turned out to be the least interesting part of the problem.
The more important question was:
Why do people ask a fortune stick a question in the first place?
The Question Behind the Question
Someone asking whether a relationship will survive may appear to be asking for a prediction.
But often there is another question underneath it:
Should I keep trying?
Someone asking about a career fortune may really be wondering:
Am I wasting my life here?
Someone asking whether the coming year will be fortunate may actually be asking:
Am I going to be okay?
A traditional fortune stick cannot know the person's complete situation.
Nor should an AI pretend that it can predict the future.
But the combination suggested a different product direction.
The fortune stick could provide the trigger: a symbol, story, or unexpected perspective that encourages reflection.
AI could provide something different: help the person examine the situation rationally, consider alternatives, calm the immediate emotional response, and identify practical next steps.
The roles should remain distinct.
The fortune stick brings intuition, symbolism and insight.
AI brings reasoning, perspective, emotional steadiness and action.
Once I looked at the product this way, I was no longer primarily designing a fortune-telling application.
I was designing a space in which someone could think.
Where Can People Still Be Completely Honest?
That led to a harder problem.
For an AI system to help someone think through a difficult situation, the person first has to tell it what is actually happening.
That is not always easy.
The most valuable conversations may involve things people hesitate to tell colleagues, friends, family members or even partners.
A relationship may be failing.
A career may have stalled.
Someone may be ashamed of a decision.
They may be frightened, jealous, angry, lonely or simply uncertain.
An AI interface creates an unusual possibility: a place where a person can articulate thoughts without immediately worrying about another person's reaction.
But it introduces another problem.
Where does the conversation go?
Most modern AI products are built around cloud inference. A message leaves the device, travels across the network, is processed by remote infrastructure and a response comes back.
Technically, this architecture is straightforward.
Psychologically, it changes the conversation.
If users believe that their most private thoughts are leaving the device, some will censor themselves.
The product can promise empathy.
The interface can say that it is safe to talk.
But architecture ultimately determines what those promises mean.
That made privacy more than a compliance requirement for 觀音AI .
Privacy became part of the product itself.
Moving the AI to the User
The alternative was technically much more difficult:
run the language model on the user's own device.
Instead of sending the conversation to a remote AI service, inference can happen locally.
This changes the architecture, but more importantly it changes the relationship between the user and the product.
The phone is no longer merely a terminal connected to an intelligence somewhere else.
The intelligence is inside the device.
Private conversations can remain there.
That decision comes with significant trade-offs.
A cloud model can be enormous. A mobile model cannot.
Cloud infrastructure can allocate substantial memory and compute to a request. An iPhone has to share limited resources with the operating system and everything else running on the device.
A cloud model can be upgraded centrally. An on-device model has to be packaged, downloaded, cached, versioned and managed.
Model size matters.
Memory matters.
Latency matters.
Battery consumption matters.
Older devices matter.
The engineering problem therefore became much more interesting than adding an API call to a chat screen.
The question was no longer:
How do I add AI to an app?
It became:
How much useful intelligence can I put directly into a person's hands without requiring that person's private conversation to leave the device?
A Fifteen-Year Product Meets a New Technology
There was another reason this project mattered to me.
觀音AI and my other Chinese cultural applications did not begin as AI products.
They came from an earlier generation of mobile software.
The original design assumptions belonged to the App Store era that began around 2008 and 2009: applications were collections of screens, data, interactions and carefully designed workflows.
Generative AI challenges that model.
Instead of navigating through a predetermined information hierarchy, users can increasingly express what they want in natural language.
Instead of merely retrieving an interpretation, they can question it.
Instead of reading a static answer, they can explain their own circumstances and continue the conversation.
This does not mean every old application needs a chatbot.
It means AI creates an opportunity to reconsider the fundamental job the product was supposed to perform.
For a fortune-stick application, that distinction is important.
If the job is simply:
"Show me the interpretation of fortune stick number 37."
then a database lookup already solves it.
AI adds unnecessary complexity.
But if the job is:
"Help me understand what this means in the context of what is happening in my life."
then the problem is completely different.
That is where AI becomes useful.
AI Should Not Replace the Fortune Stick
There was an easy product direction I deliberately did not want to take.
The AI could have become an electronic fortune teller.
Ask it a question and let the model generate a mystical prediction.
Technically, that would be simple.
Conceptually, it would destroy what made the original experience interesting.
The fortune stick and the AI should not compete for the same role.
The traditional system provides ambiguity.
That ambiguity is useful.
A poem or symbolic story does not give the user a perfectly optimized answer. It creates enough distance from the immediate problem to make another interpretation possible.
AI is useful after that moment.
It can ask what has happened.
It can distinguish facts from assumptions.
It can explore several interpretations.
It can suggest actions.
It can help turn an abstract insight into something the user can actually do.
The product therefore became an experiment in combining two very different systems of thought rather than using one to imitate the other.
The Architecture Became the Product
Building 觀音AI forced several decisions that would normally be considered implementation details.
Should inference happen in the cloud or on the device?
How large can the model realistically become?
Which devices should be supported?
What happens when there is no network connection?
What information needs to persist?
How should conversation history be stored?
How do you communicate privacy without turning the interface into a legal document?
These are architecture questions.
They are also product questions.
A privacy promise implemented only in marketing copy is weak.
A privacy promise implemented by removing the need to transmit the conversation is fundamentally different.
This became one of the central lessons of the project:
Architecture can be part of the user experience even when the user never sees the architecture.
Beyond Fortune Telling
觀音AI is still recognizably descended from a fortune-stick application.
The cultural context matters.
The ritual matters.
The symbolism matters.
But those are no longer the boundaries of the product.
The more interesting opportunity is using that familiar interaction as the beginning of a private process of reflection.
A fortune stick can introduce an unexpected perspective.
AI can help examine it.
The user still makes the decision.
That last point is important.
The goal is not to build a machine that tells people what their future will be.
It is to build technology that helps people think more clearly about what they might do next.
And if those conversations are genuinely private, people may be willing to tell the technology things they would otherwise never say.
That is where this project stopped being primarily about fortune telling.
And that is where the engineering became interesting.
End of Cover Story