Co-founder · EdTech startup · Feb – Dec 2025
LectureAI
LectureAI started with an email from the NUS School of Computing about a startup programme. Arshin, a good friend who's really into startups, asked if I wanted to do it with him, and with Vidushi we came up with the idea: take a lecture recording and turn it into proper study notes. That programme rejected us, but BLOCK71 accepted us into its incubator.
01What we built
We surveyed students across NUS first, and it was a real problem. Then we built the pipeline. FFmpeg processes the audio, it gets transcribed, technical terms in the transcript get corrected using the lecture slides, and then an LLM turns it into summaries and study notes. We fixed the transcript before summarising because when a transcript gets a technical term wrong, the summary repeats the wrong term with total confidence.
The backend was FastAPI, with Whisper for transcription and Alembic for database migrations, and the frontend was React, which ran each step in turn and showed progress as it went. Later we added a background job queue with Celery and Redis for long lectures, but the app hadn't switched over to it by the time we stopped. I led the full-stack development, and the notes it produced were really good.
02Why we stopped
The next step was getting into universities, through things like Canvas integrations or pilots with a course. That turned out to mean a lot of paperwork and a very slow process, probably two or three years before anything happened. We reached out to a lot of people and didn't hear back from many. Around the same time TurboScribe, a competitor with a lot of funding, took off with students directly, which closed off the other route.
So in December 2025 we wrapped it up. I learned a lot from it, from surveying users and pitching to building the pipeline.