Skip to content
Discuss your product
EdTech and learning platforms

An AI tutor over your own platform's content

The version almost every platform builds first is a general model with a system prompt saying "you are a tutor for this course". It works in the demo and fails in class, because the student asks about the notation that instructor uses, the example in unit three, the rubric criterion. None of that is in the model.

The version that actually works starts with the content: transcribe the video, split the material into passages with their references, and build retrieval before conversation. The tutor is the last layer, not the first — and the cheaper of the two.

Short answer

A useful AI tutor isn't a general chat with the course's name on it: it's a system that retrieves passages from your own material, answers citing the module and the timestamp, says "that isn't in this unit" when it isn't, and when facing a graded assignment guides to the step before without solving it.

How it works

01

The content gets prepared

Timestamped video transcription, splitting material into passages, and metadata for course, unit and learning objective. This is half the project's work.

02

Retrieval is scoped to the student

Only the courses they're enrolled in and the units they've already unlocked. A tutor that answers with unit seven's material wrecks the course's pedagogical sequence.

03

The answer is tied to a citation

Every claim comes with the passage supporting it and a link to the video timestamp or the page. No passage, no claim: the tutor says it isn't in the material.

04

The assessment boundary gets defined

The tutor recognizes when a question is literally the graded prompt and switches mode: it explains the concept, offers an analogous exercise, and withholds the answer. That boundary is tested against the course's real assignment texts.

What gets measured

  • Share of answers with a verifiable citation to the material.
  • Correct-abstention rate: out-of-material questions the tutor declines instead of answering.
  • Forum or instructor questions that stopped being asked — and which ones remain.
  • Model cost per active student per month.

What's needed on your side

  • Content accessible via API or exportable, and video with transcripts or a budget to produce them.
  • Knowing, per student, which courses and units are unlocked.
  • The assignment texts, so the boundary can be defined.
  • A per-student cost ceiling set by product before building.

When it isn't worth it

  • If the content is mostly untranscribed video with no budget to transcribe it, the tutor will answer from the written fraction and the student will notice on day one.
  • If the course teaches judgement rather than information — design, writing, negotiation — the value of retrieving passages is low and the project looks more like assisted grading than tutoring.
  • If the student's plan costs less per month than the estimated model cost, there's no product. That calculation comes before the prototype, not after.

Related questions

Doesn't it just become a way to cheat?
It does if it's built without the boundary. With the boundary defined and tested against the real assignment texts, the tutor explains the concept and offers an analogous exercise instead of solving the prompt. It also leaves a conversation log, which is more than exists today when the student uses a general model on their own.
What happens when the course material is outdated or has an error?
The tutor repeats it, because it answers from the source. That sounds like a defect and is a feature: the citation makes the error visible and therefore fixable, whereas a model answering from memory papers over it with a correct answer that contradicts the material without anyone noticing.
Can it launch to the whole user base at once?
You can, but it's better not to. Cost per student behaves differently under real traffic than in the estimate, and one course's question pattern doesn't resemble another's. One course first, with a spend ceiling and correct-abstention measurement, informs the decision far better than a general launch.

Use cases in this industry

Rubric-based AI assisted grading

How to implement AI assisted grading in a learning platform: a draft grade per criterion, evidence quoted from the student, and the teacher's signature.

Semantic search over learning content

How to implement semantic search in a learning platform: over transcribed video and your own material, citing the timestamp and respecting the student's scope.

Next step

What education product should exist next?

Tell us what you are building, what is not working yet or which AI opportunity you want to evaluate. In the first conversation, we will tell you where we would start, what it requires and what we would not build.

Discuss your productSee the whole industry
Message us on WhatsApp