Why I built it
I wanted to help classmates who were struggling with the Abitur. I also wanted to try out the new voice transcription models and see how well small models can grade an exam.
What I built
mockabi lets students rehearse the oral Abitur with an AI examiner in the real format, with 20 minutes to prepare, a 10-minute talk and a 10-minute discussion. You speak, the examiner’s questions show up on screen, and at the end you get a practice score from 0 to 15 points with tips. It covers German and maths in five German states. I built it on my own with Next.js, Supabase, Gemini, gpt-4o-transcribe and Stripe.
The hard part
It took a lot of rounds until the examiner acted like a real teacher. Early versions praised answers, mentioned being an AI and added stage directions. Speech was just as much work. Recordings failed, transcripts went missing and the microphone sometimes cut out mid-answer, so now the app keeps whatever was recorded until then.
What I learned
Getting the model to work was the quick part. Most of the time went into everything around it, like recovering audio, timing, cost limits and payments.
Results
mockabi is live at mockabi.de, and payments through Stripe have been on since June 2026.