Atilak Modetad

Computer Vision · 3D Geometry · Applied Machine Learning

Applied Computer Science student at KMUTT working across computer vision, 3D point-cloud geometry, and applied ML. First-authored a knowledge-distillation method presented at an international conference; currently designs and builds — largely solo — a multi-view rendering and consensus pipeline for automated heritage-site damage detection.

Publication

Knowledge Distillation Using Entropy-Guided Dynamic Parameter Transfer for Wound Image Classification

Apr 2025

1st author of 6 · ICCI 2025, International Conference on Cybernetics and Innovations · Pattaya, Thailand

  • 95.22%accuracy
  • 42.7 MB (−49%)model size
  • 2.7 ms (−36%)inference
  • +4.66 ptsvs. baseline KD
  • Designed entropy-guided temperature (EGATS) and loss-balance (EGAAS) scaling that adapts knowledge distillation to each sample's uncertainty, plus a feature-alignment term, to compress a ResNet-34 wound classifier into a ResNet-18 student.
  • Built the full pipeline (augmentation, fastai teacher, pruning and layer-truncation baselines, evaluation) and released the code and trained models.
  1. ResNet-34 teacher
  2. Entropy per sample
  3. Adaptive T and α
  4. Distillation loss
  5. ResNet-18 student
  • PyTorch
  • fastai
  • Albumentations
  • Knowledge distillation

IEEE XploreCodeFull case study →

Experience

Research Assistant — Automated Heritage Damage Detection

Dec 2025 – Present

"Ancient" project, KMUTT · grant-funded, paid in installments

Independent contribution

Overall accuracy 90.08% → 99.52% · Crack IoU → 83.25% · Spalling IoU → 97.28% · Decay IoU → 98.43%

  • Independently designed and built a multi-view 3D→2D rendering and consensus-classification pipeline, then rewrote the renderer around EWA (Elliptical Weighted Average) surface splatting to close rendering gaps at oblique camera angles.
  • Also built, independently, a point-cloud simplification pipeline cutting a real 5.49M-point heritage scan to 1.53M points (−72%) while preserving damage-relevant detail.
  1. TLS point cloud
  2. 32 camera views
  3. EWA splatting render
  4. Per-pixel consensus
  5. Damage labels
  • Python
  • Open3D
  • PyVista
  • DBSCAN / HDBSCAN
  • Google Colab GPU

Full case study →

Projects & Competitions

FahMai RAG Challenge

2026

Super AI Engineer Season 6, Level 1 Hackathon

  • 100 × 10 choicesquestions
  • 118 docs · 857 chunksknowledge base
  • 5intent routes
  • Built an intent-aware hybrid retrieval pipeline for Thai product Q&A: BGE-M3 dense search and BM25 over PyThaiNLP tokens, fused with intent-weighted Reciprocal Rank Fusion.
  • Engineered forced retrieval for every product named in a question, and prompts that answer "not in the knowledge base" instead of guessing.
  1. Question + 10 choices
  2. Intent routing
  3. BGE-M3 + BM25
  4. Fusion + forced retrieval
  5. LLM answer
  • BGE-M3
  • BM25
  • PyThaiNLP
  • RRF
  • OpenThaiGPT (ThaiLLM API)
  • Google Colab

CodeFull case study →

Thai Open-Data Storytelling

Mar 2026

Super AI Engineer Season 6, Mini-hackathon

  • 3open datasets
  • 65provinces scored
  • 2.0Mstudent records (2025)
  • Turned three Thai open datasets — fire statistics, university enrollment, rainfall and flood risk — into question-driven analyses with tiered recommendations and dashboard-ready CSVs.
  • Scored fire risk per province on a frequency × damage plane (23 of 65 critical); factories were the strongest predictor of severity not built into the score (r = 0.73).
  1. THackle open data
  2. Audit and clean
  3. Features and scores
  4. Question-driven analysis
  5. Recommendation + dashboard CSV
  • pandas
  • Plotly
  • Looker Studio-ready CSV
  • Google Colab

Code

Point-Cloud and Classical-ML Experiments

Self-study

Coursework and self-study

  • 12,311ModelNet40 meshes
  • 67.1%DGCNN, 1 epoch
  • 89.3%RBF SVM, optimistic
  • Trained a Dynamic Graph CNN from scratch on 512-point clouds sampled from ModelNet40, and repaired the dataset's malformed OFF headers so Open3D could read them.
  • Compared linear and RBF SVMs on pixel features of spider and scorpion photos, and documented why the result is optimistic: augmentation before the train/test split.
  • PyTorch
  • Open3D
  • scikit-learn
  • Albumentations

Code

Space-Hero Companion

Oct 2024

NASA International Space Apps Challenge, Space Challenge Asia regional

Galactic 2nd Runner-Up
  • Presenter · Design · Researchrole
  • 7 peopleteam
  • Presented, shaped the design concept, and researched the astronaut-wellbeing domain behind a companion-app game; a teammate built the Unity/C# implementation.

Full case study →

PestGuard AI

2024

AI + IoT Young Master Competition

Best Excellent Idea, ฿5,000
  • Idea / pitch onlynot implemented
  • Conceived and pitched, solo, a camera+AI pest/animal-detection concept alerting through Home Assistant.

Education

B.Sc. Applied Computer Science

In progress

King Mongkut's University of Technology Thonburi (KMUTT), Dept. of Mathematics, Faculty of Science

High School — Science–Mathematics, Robotics Engineering track

2025

Wat Rajabopit School, Bangkok  ·  GPAX 3.68

Skills

ML & Vision
PyTorchfastaiscikit-learnOpen3DPyVistaAlbumentationsPointNet / PointNet++DGCNNKnowledge Distillation
Retrieval & NLP
Sentence-Transformers (BGE-M3)Hybrid BM25 + Dense RetrievalPyThaiNLP
Data & Visualization
pandasNumPyMatplotlibSeabornPlotly
Systems & Tools
FastAPIGit / GitHubGoogle Colabmkdocs
Languages
PythonC++
Embedded
ArduinoRaspberry PiKidBright

Awards & Recognition

2024RAC Cohort 2, IPST / PMU-B national program — selected 1 of 25 nationally2024NASA Space Apps Challenge — Galactic 2nd Runner-Up2024ICT Challenges, KMUTT — top 10 nationally, honorable mention2024AI Prompt SPU Mini Hackathon — 4th place, honorable mention2024AI + IoT Young Master Competition — Best Excellent Idea, ฿5,0002025ICCI 2025 — first-authored paper accepted & presented