DEV Community

Cover image for Building a live AI tutor that teaches on a blackboard
Touhid
Touhid

Posted on Fully Autonomous

Building a live AI tutor that teaches on a blackboard

A recorded lesson has a fixed next step. When an explanation stops making sense, the learner has to pause, search elsewhere, or continue without understanding it. I wanted to explore a different interaction: a tutor that can explain visually and respond when the learner interrupts.

That is the idea behind MyLecture, a desktop app for Windows and Mac. Its live AI tutor speaks while writing, drawing and animating on a blackboard. Learners can ask follow-up questions, bring an image into the lesson, and answer practice checks. The app also includes a studio for narrated animated videos, but the tutoring interaction starts with a smaller goal: make the confusing step discussable.

A research motivation, with a boundary

Benjamin Bloom's 1984 paper, The 2 Sigma Problem, discussed studies comparing conventional instruction, mastery learning and tutoring. The average student in the tutoring condition scored about two standard deviations above the conventional-class average, corresponding to performance above roughly 98% of that comparison group. The tutoring condition included formative checks and corrective feedback. The original paper's first page describes the setup.

That finding concerns particular human-tutoring studies. It does not show that an AI tutor, a blackboard interface, or MyLecture produces the same result. I have not established that outcome for this app. For me, the useful design question is how software can make explanation, questions, practice and correction available in a more responsive interaction.

Why the blackboard matters to the interaction

A block of text can contain a correct explanation while still making relationships hard to see. A diagram gives the tutor something to point at. Animation can show a relationship changing over time. Writing and speech can develop an explanation together instead of asking the learner to translate between a paragraph and a separate picture.

The current tutor walkthrough uses an eclipse lesson. The explanation includes an image and a moving Sun–Moon–Earth diagram. This is a concrete example of the interaction I am building: the learner can watch the relationship being explained, then ask about a particular part of it.

Visuals are not automatically good teaching. An animation can distract, a diagram can be misleading, and spoken narration can get ahead of what is visible. The useful question is whether a learner can connect the explanation to the thing on the board. That needs observation and feedback, rather than an assumption that more movement means more understanding.

Starting as a learner

The app is free, but the connected AI service is separate. A learner connects their own ChatGPT, Claude or Gemini plan, or an API key. That provider's limits and charges apply. Running the desktop interface on a personal computer does not make the connected model offline or remove the need for a network connection.

After connecting a provider, a useful first request is a small, specific topic with some indication of what is confusing. For example: explain why an eclipse does not happen every month, using a diagram. This is a suggested learning prompt, rather than a promise about an exact scripted response.

The tutor then has a visual surface for its explanation. As a learner, I would begin with one question and watch whether the explanation addresses it, instead of requesting an entire course before checking the first concept. A short initial exchange makes it easier to notice when the explanation has taken a wrong direction.

Interrupting is part of the lesson

MyLecture supports asking a question during a lesson and receiving a fresh visual explanation. That changes how a learner can use the board: an unclear diagram can become the subject of the next exchange, rather than something to silently work around.

For instance, a learner might ask for a slower explanation of one relationship, or ask how a newly introduced term connects to the picture. These are examples of ways to use the conversation, not claims that the app detects confusion automatically. The learner still needs to signal what they want clarified.

Images provide another starting point. A learner can attach a picture and ask the tutor to explain it. A useful question identifies the region or concept of interest. A poor-quality image or an ambiguous prompt can lead to an incorrect interpretation, so the learner should check that the tutor is explaining the intended object before building on the answer.

Practice creates another opportunity to correct

The tutor can present multiple-choice questions and give feedback on an answer. That adds a moment where the learner has to choose, rather than only watch. Follow-up explanation can then address the response.

An immediate correctness label is not proof of understanding. A learner can guess correctly, and AI-generated questions or feedback can contain errors. I would use a practice check as a prompt to explain the reasoning: why does the chosen answer follow from the diagram? For important coursework, checking against trusted course material remains useful.

The next product question is whether these interactions help people learn and return. A completed lesson or an exported file alone does not answer that. Evidence about understanding needs an appropriate evaluation, and I do not want an attractive walkthrough to stand in for it.

Where the video studio fits

MyLecture also lets creators turn an idea or notes into a narrated animated MP4. The studio supports longer videos, short reels, product demos and explainers. Creators can change a voice, edit a narration line and regenerate it, choose a theme or customize one, and set the export format.

That is a different workflow from live tutoring. A video has a planned sequence that viewers can replay; a live lesson can respond to a question. Landscape, portrait and square exports, together with resolution, frame-rate and caption controls, help the creator choose how to present the finished video.

MyLecture is free for Windows and Mac: download the latest release (more at mylecture.ai). Our walkthrough videos show the real app with editorial changes, including removed waiting time; they are demonstrations of interaction, not generation-speed measurements. If you try the tutor, the most useful feedback is specific: which explanation helped, which step was confusing, and what question you asked next.


Disclosure: I'm the maker of this project. This post was drafted with AI assistance and edited for accuracy against the project's code and documentation.

Top comments (0)