← 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

A student can upload several pages of handwritten homework. LearnLoop reads the work, points out each mistake and links it back to the lecture slide that teaches the idea. Its tutor answers from approved course material and always shows the source slide.

Students launch it from SUTD's learning platform without creating another account. Under the hood, that meant LTI 1.3 with OIDC and signed JWTs, plus a three-legged OAuth sync for course content and announcements. RBAC, rate limits and per-user AI caps keep access and spending bounded.

The instructor view groups repeated mistakes across the cohort, so lecturers can see what needs reteaching without reading every student conversation.

The result

LearnLoop was piloted with 40 students across two classrooms, with a 50-student pilot confirmed for September 2026. It also received S$2,000 from SUTD's Baby Shark Fund and passed the university's security and data-protection review.

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.

Next case studyRankr