GenEd

For students, the main struggle is that courses out there are not personalised and catered for the general public. A student might excel in mathematical reasoning, but fail at explaining the concept in words, vice versa. But a course is more likely to focus on the mathematical logic rather than explanation, when sometimes what a student needs is to know how to explain their thoughts. This is where GenEd comes in, it's meant to be personalised towards a student's learning style and habits as well as preferences, making microlearning (and learning) more accessible and fun.

The generated course is done through several LLM calls, allowing for the module titles, quizzes, and content to be more predictable and let an educator or student regenerate one weak section without discarding everything else.

When an educator uploaded PDFs or supplied URLs, we chunked and embedded the material, retrieved relevant passages, and added them to the generation context. It kept domain-specific courses closer to the educator’s sources, but sadly, this didn't manage to make hallucination disappear. As a result, we made it such that every module would be passed through a human review step before publication.

We used Twilio APIs to test what GenEd could look like as a WhatsApp microlearning bot. These English, Chinese, and German runs returned useful links and small next steps without asking the learner to open another platform. It works in other languages too, and was pretty successful as a first pass. This caught the eye of a collaborator who was backed by the Malaysian Government, and allowed us to receive an initial Letter of Intent.

It was a tough decision to make, but when SUTD’s VIE office closed the funding window, the three of us decided not to seek more capital. The product worked; but making it reliable enough for real courses was a much larger job. We took the experience with us instead of stretching the project past the point where the team wanted to keep going. The project continues to live through SUTD's undergraduate research project, where we explore more AI-driven PDF-to-course generation, since this was our main bottleneck during beta testing.