From banking cores to side projects — built to ship and to last.
From banking cores to side projects — built to ship and to last.
Products and games I made, tutorials, AI models, enterprise solutions and research. Pick a category to explore.
The former presentation section now leads to the thesis category. The technical leadership interview presentation is not displayed; its existing document link is preserved.
An AI content platform with admin governance and an Event Radar; it curates security and AI news every day.
Stock health checks, tracking and notifications — five data angles, risk first, a daily brief.
An online question bank organized by exam and subject for practice, revision and certification preparation. Authorized login required.
A 3D board, an AI opponent and a step-by-step tutor that adapts to your level.
Play five-in-a-row online against the computer.
No Blender needed — from reference art and an AI first pass to a presentable static model, in a 19-slide course deck.
Read on site
An eight-week guide using insurance content on IG/Facebook: audience, topics, scripts, filming, editing, publishing and reviewing audience response.
Read on site
This study asks a model to find plastic bottles, metal cans and construction debris in mixed-waste photos and draw a box around each object. Images and human labels are prepared first, then YOLO26n is trained to compare image sizes, unwanted boxes and video processing speed.
On the same 200 test images, the model trained with larger inputs gains 4.48 points. The score is mAP50–95, shown on a 0–100 scale. It assesses object labels and box placement; 55.38 points does not mean 55.38% accuracy in the field.
Human review found no clearly identifiable targets of the three classes in these photos. Any predicted box is therefore a false positive. B2 produces 29 unwanted boxes instead of 76, but affects 16 photos instead of 15. Fewer boxes does not yet establish reliable field detection.
The work establishes a data-to-training-to-review workflow and measures the trade-off between detail and speed. The next step is to collect field photos with clearly visible targets, so missed bottles, cans and debris can be measured as well.
Read on site
An unofficial replica of a corporate website with a data-center back office, deployed on Cloudflare.
A discussion draft for adopting AI tools in a company: planning (4W1H), usage & governance rules, an AI-assisted development SOP, an install & operation manual, and templates.
Read on site
A brand website bringing bird care, pet products and store information together.
Shift scheduling and attendance for an enterprise HR system.
Built with Academia Sinica; live at Linkou Chang Gung and Taichung Veterans General, among others.
A two-stage trading optimization framework built on contrastive learning and knowledge distillation (ETF 00891).
Read on site