Connecting scholarship notices to an AI assistant, how we did it

The problem

Scholarship notices were scattered across offices and inconsistent formats, so students had to dig through memos just to find out what they could apply for, while staff answered the same questions over and over by hand. That was exactly the problem we ran into building the scholarship platform for NCUE's Office of Student Affairs: the notices themselves were never structured in a way an AI could read, so any assistant built on top of them could only handle surface-level questions and stumbled on the details.

How we broke it down

We split "make the AI answer correctly" into a few concrete engineering problems, instead of leaving it to the model to figure out:

• Centralize the notices before adding AI. Every on- and off-campus scholarship notice lives in one platform with the same fields: eligibility, deadline, amount. Skip this step and whatever AI you bolt on later only ever learns from incomplete information.
• The data lifecycle is built for AI. The moment a notice is created or updated, it's rewritten into a form the AI can search and stored in the knowledge base; unpublishing removes it, and a daily job reconciles the rest so the AI never answers from stale data. The assistant always searches the knowledge base first and only goes online, citing its source, when nothing is found.
• Memory asks before it remembers. The AI can suggest remembering things like a student's status, grades or residence so they don't have to repeat themselves next time, but nothing is written until the student taps "agree." The memory itself can be viewed, edited or cleared.
• One account, three interfaces. Web, the Android app and the LINE official account sign in to the same account and see the same data. A deadline reminder set on the web also arrives in the app and on LINE, and a conversation with the AI carries over between web and LINE without repeating the context.

None of these pieces is hard on its own. What's hard is wiring them into one line: every change to a notice has to correctly propagate to what the AI can see, instead of maintaining a separate "AI-only" copy of the content that quietly drifts out of sync.

The result

Once the platform was live, the AI answered directly from the notices, and the office no longer had to reply to common questions one by one. By the student affairs office's own estimate, the time spent handling these inquiries dropped by 98%. Conversations students thumb down also feed a knowledge-gap analysis, which the AI turns into draft FAQ entries for the office to review and publish — so the answers keep improving instead of staying frozen at launch-day quality.

What we learned

An AI assistant's quality mostly comes down to how clean and current the data feeding it is, not how strong the underlying model is — getting the data lifecycle right usually beats swapping in a bigger model. "The AI decides for you" and "the AI prepares it, you confirm" are two different things, and the latter is the right shape for anything touching a student's personal data, like reminders or memory. Sharing one account and one dataset across interfaces looks like a convenience, but it's really the precondition for the AI giving the same answer no matter where you ask — skip it, and each platform ends up telling its own story.