About MelbMind

About MelbMind

MelbMind started from a simple question: what if an AI study tool understood not only the text inside course documents, but also the academic relationships between lectures, tutorials, assessments and official solutions?

It began as a retrieval prototype that answered questions from course PDFs. Over time it became an academic intelligence and Agent system: it models how a course is put together, links each assessment to its official solution, and shows the evidence and steps behind every answer.

Why I built it

Many AI study tools are built to produce an answer. While building MelbMind, I kept asking a different set of questions:

• Where did this answer come from?
• Does the evidence match what the student actually asked for — this year, this assessment, this topic?
• Is this exam question paired with its own official solution, not another paper's?
• Did the Agent actually finish the task, or only produce text?
• What should happen when the evidence is not enough?

Taking these questions seriously shaped the design. MelbMind models the structure of a course and chooses evidence by the role each document plays. Before searching, it writes down what a request really asks for (the Query Contract). Before an answer can count as complete, it checks that the evidence fits that scope (the Evidence Gate). And it records every step, so you can see how an answer was reached.

What I learned

Building a reliable Agent is not mainly about giving the model more tools. Failures need to become observable, constrained and testable.

1. What I saw

   A run reported success with no evidence behind it: no tool calls, no citations.

   What changed

   Completion now belongs to the Agent loop: a course question counts as complete only when its answer is attached to retrieved evidence. Otherwise the run ends with a named reason.

2. What I saw

   A follow-up such as “only from 2024” did not stick, and questions from outside that scope were cited as if they fitted.

   What changed

   Every run resolves a Query Contract with hard constraints, and an Evidence Gate drops anything out of scope. If nothing fits, the run says so.

3. What I saw

   A model kept repeating the same planning step: ten provider calls, seven of them invisible in the trace.

   What changed

   The loop is bounded, and every decision is either accepted or visibly rejected with its reason, so a stall is caught instead of hidden.

4. What I saw

   Matching solutions by year and question number alone could pair a question with another paper's solution.

   What changed

   Exam papers and official solutions are now linked explicitly — automatically only when a pair is unambiguous — and questions are resolved inside the linked document.

5. What I saw

   Models returned empty or truncated structured output, and providers sometimes timed out.

   What changed

   These failures are classified, retried once with a changed request, and then reported as a clear terminal reason instead of failing silently.

Each lesson ended as a test and a named terminal reason, and the evaluation runs reproducibly on the synthetic course, so a fix in one layer cannot quietly break another.

The public demo and the full version

This public demo is based on MelbMind's stable v1.0 core and is intentionally restricted for safe public access.

The public demo uses SAMPLE101, a fully synthetic course. Real university course materials are intentionally excluded: lecture slides, tutorials, assessment papers and official solutions may be protected by copyright, licensing or distribution restrictions.

The public demo is intentionally read-only. To work with your own course materials, run the full self-hosted MelbMind — your files stay in your own environment.

This is a deliberate choice: the public demo is synthetic, reproducible and safe to share, while in self-hosted MelbMind your own course materials stay in your own environment.

See the public demo and the full version side by side

Open-source release prepared. The public repository will be available after the competition. AGPL-3.0-or-later

Roadmap

Directions rather than promises, and without dates.

Project status

Public demo

SAMPLE101 synthetic environment, based on the MelbMind v1.0 core

Project

MelbMind, created by Eric (Ziyang Zhang)

Licence

AGPL-3.0-or-later

Source code

Public release after the competition

Demo course

SAMPLE101 — fully synthetic

Contact

MelbMind is built and maintained by Eric (Ziyang Zhang). Questions, feedback and collaboration ideas are welcome.