Built an AI Study Tool That Refuses to Give You the Answers for UoM

This is a submission for the DEV Weekend Challenge: Community

The Community

I am an IT undergraduate at the University of Moratuwa(UoM), Sri Lanka. In our community, the night before a brutal technical exam is chaotic. When faced with a massive, complicated past-paper question that we don't understand, students often hit a wall of panic and just give up (we call it going "Bora").

The problem with existing AI study tools is that they just give you the final answer or the raw code. That doesn't help you learn; it just helps you cheat. I wanted to build a tool for my batchmates that acts like a senior student holding a late-night study session(a "Kuppi") guiding you to the answer without actually giving it to you.

What I Built

I built Triage AI, a cross-platform (Next.js Web + Flutter Android) tool designed specifically for technical exam preparation. You simply snap a photo of a difficult past paper question, and the app uses AI to break it down.

Instead of solving the problem, it returns three specific things:

Core Concepts: The exact fundamental topics you need to Google to understand the question.

The Trap: It identifies the subtle pitfall or trick hidden in the question designed to make students fail.

Attack Plan: A step-by-step strategic breakdown of how to approach the solution.

If you are still stuck, Triage AI features a "Practice" mode that generates a similar question, and a "Hint" system that provides progressive, helpful nudges (never the answer) to get you unblocked.

Demo

Live: Vercel
This weekend sprint proved to me that AI doesn't have to be a shortcut that ruins learning. When built with the right constraints, it can be the ultimate teaching assistant.

Code

Source Code: Github

How I Built It

Balancing this weekend challenge with my strict university workload was tough. As an IT undergraduate, I couldn't afford to ignore my own university assignments to build an app. Because I was juggling coursework and only had the weekend to code, I needed to keep the technical setup extremely lean and fast.

The Brain: I used the Google Gemini 2.5 Flash API because it can process uploaded images incredibly fast.

The Logic: I built a Next.js application to handle the requests. I wrote strict prompts and set up Gemini to return structured data, which made it easy to display the results in clean UI cards.

The App: The Next.js backend powers both a smooth React web app and a native Flutter Android app(kinda buggy though) at the same time.

Aside from time management, the hardest technical challenge was the Hint System. To ensure the AI didn't repeat itself or accidentally reveal the answer, the app keeps track of the clues you've already seen. It sends this history back to Gemini so that the next hint is a progressively more specific nudge in the right direction.

Good luck with exams, batchmates!