← All work

AI study infrastructure

LearnLoop

An AI study layer inside SUTD's learning platform. It answers from approved course material and keeps the exact source slide beside every response.

40 studentspiloted across two classrooms

Passed SUTD EduTech's first cybersecurity review by moving students behind LTI 1.3 and EASE SSO, removing public signup and reducing Blackboard access to least privilege.

What it does

Course material at SUTD lives across slides, assignments and the LMS. LearnLoop adds an AI study layer inside that workflow. It retrieves from approved course material, answers beside the source slide and supports both Socratic tutoring and handwritten submissions.

The result

The product was piloted with 40 students across two classrooms and received S$2,000 from SUTD's Baby Shark Fund. It passed SUTD EduTech's first cybersecurity review after I redesigned who could access it and what the LMS integration was allowed to read.

The security decision that changed the product

The quoted penetration test cost more than the project's funding. I did not weaken the review. I moved students and instructors behind LTI 1.3 and the university's EASE login, removed public signup and reduced Blackboard access to least privilege. That changed the trust boundary enough for the S$7,000 test to be waived.

Citations are still part of the response contract, not decoration added after generation. Retrieval never ignores the course boundary. A strong semantic match from another class is still the wrong answer.

What changed

The first version tried to do too much. The useful version became narrower: answer from the course, show the evidence and stay out of the lecturer's way. That constraint made the product easier to trust and easier to fit into the LMS.

What I learned

The chat box was the easy part. Shipping inside a university meant authentication, privacy and least-privilege access mattered more than another model feature.

Next case studyRankr