Applied Computer Science student at KMUTT
Computer vision · 3D point clouds · applied machine learning
Portfolio · LinkedIn · atilak.work@gmail.com
- Published. First author of an ICCI 2025 paper on entropy-guided knowledge distillation for wound-image classification (IEEE Xplore).
- Now. Research assistant at KMUTT, designing a multi-view rendering and consensus pipeline that detects damage on heritage sites from laser-scan point clouds (private research code).
- How I write. Every project below reports its results as logged, with the limits of the evaluation next to them.
| Project | What it does | Highlights |
|---|---|---|
| Entropy-guided KD | Code for the ICCI 2025 paper: a ResNet-18 learns from a ResNet-34 whose own uncertainty sets the temperature and loss weight per sample | 95.22% against 90.89% for standard KD on 9 wound classes · 49% smaller student · evaluation caveats documented |
| FahMai RAG | Thai product Q&A over a Markdown knowledge base, 100 multiple-choice questions (Super AI Engineer S6) | Rule-based intent routing · BGE-M3 + BM25 fused with RRF · forced retrieval of named products |
| Thai data storytelling | Three open-data notebooks that end in a recommendation: fire risk, university enrollment, rain and floods | Severity scoring · 2024 vs 2025 enrollment shifts · dashboard-ready CSVs |
| ML learning notes | SVM on spider vs scorpion photos; DGCNN on ModelNet40 point clouds | Baselines with their mistakes written down |
| Deep learning & vision |     |
| Retrieval & NLP |    |
| Data & ML |     |
| Workflow |   |
Each repository has an MIT licence, a requirements file, Open-in-Colab links, a pipeline diagram and notes on the limits of its results. The paper's model weights ship with SHA-256 checksums.