Skip to content
JNP
All projects

Case study · Thesis project · 2026

TulongGuro

AI-Assisted Grading and Classroom Management System

A thesis project for the Computer Science program at Holy Angel University. TulongGuro is a learning management system for teachers in which a vision-language model reads student work, including photos of handwritten essays, against the teacher's rubric and drafts a score and feedback. The teacher reviews, edits and approves every result before a student sees it.

tulong-guro.vercel.app
TulongGuro landing page with the headline “Grading that gives you your evenings back.”

The problem

Large classes and paper-based workloads make timely, personal feedback hard for Philippine public school teachers, and most learning management systems only record grades. TulongGuro drafts rubric-based feedback on Grade 6 English work so the teacher can spend their time reviewing it instead of writing it from scratch.

What it does

  • AI-assisted checking: the model reads handwritten or typed submissions and drafts rubric-aligned scores, strengths, areas for growth and next steps.
  • Human-in-the-loop review: teachers validate and can edit every AI draft, and releasing grades to students is a separate step.
  • Separate Teacher, Student and Admin interfaces; school admins manage sections, course shells and student rosters.
  • Progress dashboards that track mastery of curriculum competencies across activities.
  • Badges that students earn from performance and milestones.
  • Delivered as a progressive web app, so it runs on desktop and phones without an app store.

Engineering decisions

The AI drafts; the teacher decides
Validation and release are separate acts, so a grade reaches a learner only after the teacher has reviewed the set and published it. The thesis abstract reports that raw AI drafts matched teacher-set performance bands in only about a quarter of cases, which is why the system is positioned as decision support and not as an automatic grader.
Images are prepared on the server before the model sees them
Photographed or uploaded work is optimised server-side, then read and scored against the rubric in a single model call.
A fallback model keeps checking available
Checking rotates between a primary and a fallback Gemini Flash model, so a class is read to one standard whichever one answers.

My contribution

I co-developed TulongGuro in a four-member thesis team: a learning management system in which a Gemini vision-language model reads typed or handwritten student work against the teacher's rubric and drafts scores and feedback. I kept teachers in control with a human-in-the-loop review flow, where every AI draft is validated and editable and releasing grades to students is a separate step, and delivered separate Teacher, Student and Admin interfaces as a progressive web app, with progress dashboards for curriculum competencies and student badges.

Technologies

  • React 19
  • Vite
  • Tailwind CSS 4
  • React Router
  • Node.js
  • Express 5
  • Prisma
  • PostgreSQL (Supabase)
  • Supabase Storage
  • Google Gemini (Generative AI SDK)
  • PWA
  • Vercel & Render