How MoneySpark works
MoneySpark is a Hong Kong credit-card recommendation AI agent: every answer is grounded in a live retrieval snapshot of the card database — never written from the model's memory — and when it cannot answer, it tells you plainly.
Content and figures below are as of 2026-10-16 and updated as the product evolves.
The journey of one question
You speak plainly; the AI understands and looks things up. Answers stay attached to the real card database.
Answers are looked up before they are spoken
Card data only counts when it comes from a live lookup; when the lookup cannot support an answer, we say so instead of filling gaps from memory.
Card names, reward rates, annual fees and welcome offers always come from a retrieval result — the model has no channel to invent these numbers. As of 2026-10-16, the library holds 308 Hong Kong credit cards, growing as the database is updated.
Material reading follows the same rule: it states facts only from the policy illustration, fund KFS or statement text you pasted in this time. Fields it cannot find are stated plainly — never filled in from training memory.
What the AI does, and what the data does
The AI model only understands your question and organises the language. Looking up numbers is done by the system, and only the returned snapshot counts as card data. This rule is enforced by the system itself, not by convention.
When it cannot answer, it says so
When the system is busy, errors, or an answer fails validation, you see an empty card with a plain “cannot answer right now” — never a guessed recommendation.
Honest failure is a designed, normal path — not an exception branch.
Where your data lives
The card library and your personal data are separated by design.
The card library is read-only
The system can only read the card library — no operation can modify card data. That boundary is drawn into the architecture, not left to convention.
Your things are stored apart
Your account, conversations, memories and the material you provide live completely separately from the card library; deleting your account removes everything on your side.
Everything is visible and deletable
Your profile, memories, conversations and the material you provide can be viewed and deleted at any time; guest conversations are erased automatically 7 days after creation, with no renewal.
Fast — which is why it can be free
The cost saved by engineering is what makes the product free for users.
Within a conversation, the system reuses the context it has already processed, so follow-ups get faster — and each question stays cheap enough that the product can be completely free for users.
What you see live, what replays after a refresh, and what the model actually produced are the same content — the process is transparent, with no quiet edits on replay.
Every change passes automated tests before it ships, and answer quality is evaluated on an ongoing basis.
That's the architecture. The fastest way to check it is to ask it something.
No sign-up needed. Every recommendation can be checked against the card snapshot, and when retrieval fails it says so.