Background

I initially proposes this project in 2024 but it wasn't taken then. Meanwhile, the landscape of freely available LLMs and related tools as advances impressively. I think this is now a very suitable project for the 2025 FYP cohort.

There are already many AI-assisted online whiteboard tools. For example Lucidspark, jeda.ai, whiteboard.ai and many more. However, when used as a classroom teaching tool they have at least one of the following downsides:

• focussed on team collaboration or brainstorming rather than on instant illustrations
• feature an interface with toolbar and multiple modes which is not suitable for live teaching
• cannot be embedded into canvas or other web apps
• require a paid subscription

Requirements

UCC uses the canvas learning management system for course delivery. As far as we are aware, there is currently no supported method to integrate a whiteboard which can be used in the lecture theatre (with projector) which is versatile, easy to use and delivers lucid drawings and illustrations.

Therefore, an idea was born for a system with the following goals:

• The frontend must be embeddable without any constraints, ideally via an iframe.
• User authentication (login) should be optional.
• There should be a configurable way to use persistent storage for authenticated users, e.g. GoogleDrive, local storage etc.
• For authenticated users, the state of the trained model should be saved to persistent storage between sessions.
• Only 2D black on white drawings have to be supported.

Concepts and Ideas

Suppose a lecturer embeds this tool in a canvas page and uses it for classroom teaching. S/he draws a vertically stretched s-shape. From the accessible context (e.g. cookies, storage, history) the tool guesses a possible mathematical meaning. So the currently drawn curve might be meant as an integral sign as part of a mathematical formula. Thus, the curve "snaps" into an integral sign. However, this guess might be wrong and the teacher keeps "shaping" this curve (maybe similar to Bezier curve objects in classic drawing tools). The AI-assistant should take this retouching step as a negative reward and thus gradually adjust to the intended category of illustrations.

There are many more ways how the AI-assistant could support the creation process. For example, labels could snap into a suitable font shape and to proper positions etc.

We feel that an implementation with some proven reinforcement learning system is suitable for this project. Among other things, the bitmap input might not need much preprocessing, apart form rescaling to a predefined pixel raster.