Donmark

AI that teaches you not just answers you. Lessons, quizzes, exams and verified certificates across different subjects

Comment

The UNCP WiFi is not allowing access to our domain. Please use your phone hotspot or a different network

Inspiration

We kept watching students especially in under-resourced schools turn to ChatGPT or Google to study. They'd type a question, get a wall of text back, and call it studying. No structure, no follow-up, no accountability, no proof of what was learned. The next day they'd do the same thing from scratch.

We asked ourselves: what if the AI was actually the teacher not just the answer machine?

That question became Donmark.

What it does

Donmark is an autonomous AI education platform that delivers structured, personalised learning across 5 academic subjects. Every student gets:

  • AI-generated lesson slides powered by Anthropic Claude unique every session, tailored to their subject and grade level
  • Human voice narration via ElevenLabs every slide read aloud like a real teacher
  • Push-to-talk AI tutor hold a button mid-lesson, ask anything, hear the answer spoken back in real time
  • AI tutor gate Claude speaks a comment after every slide and the student cannot advance until it finishes, enforcing real engagement
  • Adaptive quizzes 5 questions generated by Claude targeting exactly where each student struggled
  • 30-question timed final exams auto-graded, with a verified certificate issued on passing
  • Verified certificates unique Certificate ID, publicly verifiable at donmark.org/verify.html
  • AI note scanner Gemini Vision transcribes photographs of handwritten notes instantly
  • Rich notes editor Claude summarisation, flashcard generation, ElevenLabs read-aloud, and DALL-E 3 image generation
  • Pomodoro study timer with session logging and streak tracking
  • Live progress dashboard showing lessons, quiz scores, study time, and streaks

How we built it

Backend: FastAPI (Python) deployed on DigitalOcean App Platform. All AI calls are handled server-side so API keys stay private.

Frontend: Vanilla HTML, CSS, and JavaScript no frameworks. Built from scratch during the hackathon with a stark black-and-white design.

Database & Auth: Supabase handles all student data, progress tracking, exam results, notes, and authentication.

Tool Role
Anthropic Claude Lessons, quizzes, exams, live AI tutor, note summarisation and flashcards
ElevenLabs Voice narration for every slide, tutor spoken responses, notes read-aloud
Google Gemini Vision Handwritten note scanning and transcription
OpenAI DALL-E 3 AI image generation inside the notes editor
Supabase Database, authentication, and real-time data
DigitalOcean Backend hosting via App Platform

-

Challenges we ran into

Voice synchronisation getting ElevenLabs narration to sequence correctly with the AI tutor gate was harder than expected. The tutor had to finish speaking before the student could advance, but audio playback timing across browsers was inconsistent. We solved it with a careful promise chain and fallback timeouts.

Duplicate study time logging a beforeunload event handler was accidentally duplicated three times during development, causing study time to be logged 3× on every page exit. We caught it in testing and fixed it with a lessonTimeLogged flag to guarantee it fires at most once per session.

Keeping lessons truly unique early Claude prompts would sometimes produce similar slide structures across sessions. We fixed this by injecting randomised topic angles and explicit variation instructions into every lesson prompt.

Scope under 24 hours deciding what to cut was the hardest challenge. We prioritised ruthlessly and shipped a focused, polished core over a bloated unfinished one.

Accomplishments that we're proud of

  • Built a full end-to-end AI teaching loop from lesson generation to voice narration to adaptive quizzes to verified certificates in 24 hours
  • The push-to-talk tutor works seamlessly: students speak, Claude thinks, ElevenLabs responds aloud, all in real time
  • Every certificate has a unique verifiable ID students can prove what they earned
  • The platform looks and feels production-ready not a hackathon prototype
  • Achieved a 100% success rate on the hackathon evaluation criteria

What we learned

  • Claude is a genuinely powerful curriculum engine when prompted with structure and constraints it can teach, not just answer
  • ElevenLabs voice makes a bigger difference to the learning experience than we expected engagement goes up when students hear a human voice
  • Supabase is fast for a hackathon but RLS policies need careful setup or silent update failures will catch you at the worst time
  • Vanilla JS with good architecture is faster to build and debug than a framework under time pressure
  • Scope control is a skill cutting the right features at the right time is what gets you to a finished product

What's next for Donmark

  • Expand to 200+ subjects from STEM to languages to vocational skills, making Donmark a full curriculum platform for any student anywhere
  • Teacher dashboard let educators assign subjects, track class progress, and review individual student performance
  • Mobile app native iOS and Android so students can learn on any device, anywhere, including offline lesson caching
  • Multilingual support lessons and voice narration in multiple languages starting with Zulu, Afrikaans, French, and Swahili
  • AI report cards detailed performance summaries generated by Claude after every exam, with personalised study recommendations
  • Parent portal progress visibility for parents and guardians without requiring them to log in as the student
  • School partnerships bring Donmark into classrooms as a supplementary AI teaching tool aligned to national curricula

Built With

Share this project:

Updates