CU Get Reg
Open-source course planner used across Chula. I did the backoffice for the reviews and elective data.
- NestJS
- GraphQL
- Prisma
Some of these are things other people used. Some are things I just went looking into.
Sites and apps other people actually used, mostly student-facing — a course planner, freshman-event registrations, a couple of hackathon wins.
Open-source course planner used across Chula. I did the backoffice for the reviews and elective data.
Ten days making withdrawal and resignation paperwork less miserable. Backend's mine; it won the hackathon.
The open-house site for my faculty. I built the workshop booking part, which had to survive a lot of students hitting it at the same time.
The site came down after the event.
Freshman welcome events for the university. I built the registration module — house selection and the QR-based activities.
An interactive exhibition site that a few thousand people went through. Mostly an excuse to spend real time on motion.
A job-matching site for young people, thrown together over a competition weekend. I did the front end and the OCR that reads documents off the page. It took first place.
An app that teaches massage technique by watching through a camera and a multimodal model. We placed 2nd runner-up out of more than 700 teams.
A predictive-maintenance system that forecasts equipment sensor readings to catch failures before they happen. 2nd runner-up among 156 teams.
A service that ran the Thai university admission ranking ahead of time so students could see where their scores would land them. I architected it; it served more than forty thousand people.
Coursework and research I did for class and for myself, mostly notebooks rather than websites — a Thai benchmark, a diffusion model for summarisation, chess read straight off video.
Built a Thai benchmark for multimodal models, generating the evaluation questions from Visual Genome data. Mostly a lesson in how much of benchmarking is arguing about what counts as a correct answer.
Supervised fine-tuning of the LLaDA diffusion model for Thai text summarisation, evaluated against the usual suspects with DeepEval. It held up better than I expected on some things and worse on others.
A computer-vision pipeline that turns chess video into notation, fine-tuning ResNet and InceptionV3. It works well on the boards it was trained on, which is the honest caveat.
Forecasting S&P 500 returns with a handful of ML models, then running portfolio optimisation on top. The backtest looked great, which is exactly when you should be suspicious of a backtest.
A school research project on whether multimedia actually helps people take better notes. Mask R-CNN and a random forest, and a lot of trial and error.
A maths model for arguing about the greatest tennis player of all time, built for IMMC 2021. Settles nothing, which is the point.