MarkMentor

Inspiration

Every student knows the feeling: you work on an assignment for weeks, submit it, and only discover what went wrong when the grade comes back – too late to fix it. Meanwhile, AI chatbots will happily rewrite your essay, but they never look at your lecturer's actual marking rubric, and using them that way teaches you nothing. We wanted an AI that helps students learn, not one that does the work for them.

What it does

MarkMentor gives students rubric-based feedback – and keeps them the author.

Marker Agent rates the draft against each criterion of the lecturer's real rubric (Strong / Okay / Weak), quotes the student's own words as evidence, and explains what's missing.
Evidence verification – our code checks that every quote genuinely exists in the document. Unverifiable feedback is flagged for human review.
Learning Coach Agent asks Socratic questions instead of giving answers.
Academic Integrity Agent screens every AI response and blocks rewritten paragraphs, model answers or code. Ask it to "rewrite my introduction" and it refuses with a guiding question.
Feedback-Quality Agent scores how reliable our own feedback is, using automated checks – kept separate from the student's mark.
Students revise their own draft, resubmit, and watch ratings move from red to green. Progress is tracked from Assignment 1 to Assignment 2, comparing only equivalent criteria.
Lecturers import up to six code and six report questions from CSV/JSON, and see anonymised class insights (with consent and minimum-cohort privacy thresholds) plus a suggested teaching focus.
Human-in-the-loop: students can escalate flagged feedback to their lecturer, who rates each criterion 1–5 with an optional 10-word comment.
Accessibility built in: dark mode, colour-blind palettes, read-aloud, adjustable text size and reduced motion.

How we built it

Frontend: React, TypeScript, Vite, Tailwind CSS, Recharts, Lucide icons
Backend: Python, FastAPI, Pydantic, SQLAlchemy, SQLite
AI: OpenAI API in structured JSON mode, validated with Pydantic, with each agent in its own module
Pipeline: upload → text extraction (PDF/DOCX/TXT) → rubric validation → Marker → quote verification → Integrity → Feedback-Quality → save → coach → reassess
Resilience: if the AI is unavailable, a clearly labelled offline simulator keeps the demo working
33 automated tests covering guardrails, permissions, privacy thresholds and the two-round improvement flow

Challenges we ran into

Stopping AI from inventing evidence. Models sometimes paraphrase or invent quotes, so we built programmatic verification with normalised and fuzzy matching instead of trusting the model.
Guardrails that actually hold. Blocking ghost-writing needed layers: request screening, rule checks for suggested wording, code and paraphrases of the student's own sentences, then an AI classifier. It's not impossible to bypass, and we say so honestly.
Not letting the AI grade itself. Our feedback-quality score comes from deterministic checks, not the model's opinion.
Comparing different assignments fairly. Rubrics change between assignments, so we only compare criteria that assess the same skill – and tell the student when a comparison isn't possible.
Privacy. Giving lecturers useful class insights without exposing individuals.

Accomplishments that we're proud of

A full working product, not a mock-up: real uploads, real reassessment, nothing hard-coded.
The moment the AI refuses to write the essay and asks a question instead.
Every piece of feedback is traceable to a rubric criterion and a verified quote.
A guided demo that shows a weak draft improving from red to green in under three minutes.
Accessibility and privacy designed in from the start.

What we learned

Trustworthy AI in education is less about the model and more about the checks around it.
Guardrails need to be layered – and it's important to be honest about their limits.
Good feedback asks questions; students improve more when they make the changes themselves.
Lecturers value aggregated, anonymised insight more than more data about individuals.

What's next for MarkMentor

Integration with university platforms such as Moodle and Canvas, with single sign-on
A pilot with real modules to measure whether students' grades and confidence improve
Lecturer calibration: comparing AI ratings with lecturer ratings to improve accuracy
Support for more assignment types: lab reports, presentations and code with automated test runs
Multilingual feedback for international students