Structured study notes generated in under a minute

How Apoliums approached it.

Built an AI notes generator: a tutor submits a topic, structured notes are emailed to students in under a minute.

1. Constrain the output before improving the prompt

   Notes are generated against a strict JSON schema, objective, prerequisites, sections, worked example, common mistakes. A response that does not validate is regenerated once, then fails visibly instead of shipping a malformed document.

2. Ground generation in the institute's own material

   Topic requests retrieve the institute's existing syllabus text and past notes, and the model is instructed to work from that context. Ungrounded claims are the failure mode this removes.

3. Keep a human in the loop by default

   Generated notes land in a tutor review queue, not a student inbox. The tutor edits and approves; approval is what triggers the send. The queue also captures what was edited, which is the signal for improving prompts.

4. Cache by topic fingerprint

   A hash of topic, level and syllabus version keys the cache. Two tutors asking for the same topic on the same syllabus get the same approved notes rather than two generations and two bills.

What the system runs on.

• Next.js
• Python
• OpenAI API
• AWS

Front end

Next.js. A tutor console for requests and review, and a student-facing archive of approved notes.

Generation service

Python worker that assembles retrieved context, calls the model with a hard token ceiling, validates the response against a JSON schema, and retries once on a schema failure.

Retrieval

Syllabus and past notes are chunked and embedded once per syllabus version; retrieval is filtered by subject and level before the semantic search runs.

Queue

Requests are queued rather than handled in the request cycle, so a slow model response never holds an HTTP connection open.

Delivery

AWS. Approved notes render to PDF and email on approval, with per-request cost and token usage logged against the tutor and topic.

What changed after launch.

Hours of manual note-writing removed per week.

• Notes arrive in one structure across every subject, because the structure is a schema rather than a house style.

• Nothing reaches a student without a tutor approving it, so the AI drafts and a person signs.

• Repeat topics are served from cache, which makes per-term generation cost a function of unique topics rather than of tutor headcount.

• Every generation logs its token usage and its retrieved context, so a bad note can be traced to the material it was built from.