AI & ML interests

Multimodal AI for Hospitality | Siamese Networks for Image & Text Similarity | DINOv2-based Multi-head Classification | Spectral Clustering | KLUE BERT Fine-tuning | Production ML Deployment (ONNX, SageMaker) | Computer Vision + NLP Integration | Korean Language Processing

Recent Activity

jadewonΒ  updated a Space about 2 months ago
ondame/README
justin-ondaΒ  updated a model about 2 months ago
ondame/room-name-similarity
justin-ondaΒ  updated a model about 2 months ago
ondame/room-image-similarity
View all activity

ONDA - AI-Powered Hospitality Technology

🏒 About Us

ONDA is a hospitality technology company based in Seoul, applying machine learning to solve real-world challenges in hotel content management. We serve 10,000+ properties across Asia-Pacific, processing 5.4M+ annual bookings.

πŸ€– Open Source ML Research

We're developing multimodal AI systems for automated hospitality content organization and classification.

Multimodal Room Clustering

Automatically grouping hotel room images and text descriptions using vision and language models.

Image Similarity

  • Siamese network with EfficientNetV2-S backbone
  • Trained on 2M+ image pairs
  • Spectral clustering for automated grouping
  • Deployed via ONNX for production inference

Text Similarity

  • KLUE BERT-based Siamese network for Korean room names
  • Fine-tuned on 31,452 training pairs
  • Handles Korean-English multilingual text

DINOv2 Multi-head Classification

Scene classification system for hotel images achieving 85.6% accuracy.

Architecture

  • Vision Transformer (DINOv2) backbone
  • Three classification heads: Scene (6 classes), Concept (3 classes), Object (12 classes)
  • Trained on 33,762 images from 127 properties

Performance

  • Scene: 85.2% | Concept: 92.8% | Object: 78.9%

Production ML Pipeline

  • Model Optimization: PyTorch β†’ ONNX conversion
  • Deployment: AWS SageMaker for scalable inference
  • Frameworks: PyTorch, Transformers (Hugging Face), timm, scikit-learn

πŸ”¬ Research Interests

  • Contrastive learning for visual similarity
  • Self-supervised vision models (DINO, DINOv2)
  • Multilingual NLP (Korean, English, Japanese)
  • Efficient model deployment for production systems

🌏 Learn More

Website: global.onda.me


ONDA - AI 기반 μˆ™λ°• 기술

🏒 μ†Œκ°œ

ONDAλŠ” μ„œμšΈ 기반의 μˆ™λ°• 기술 νšŒμ‚¬λ‘œ, ν˜Έν…” μ½˜ν…μΈ  κ΄€λ¦¬μ˜ μ‹€μ§ˆμ μΈ 문제λ₯Ό λ¨Έμ‹ λŸ¬λ‹μœΌλ‘œ ν•΄κ²°ν•˜κ³  μžˆμŠ΅λ‹ˆλ‹€. μ•„μ‹œμ•„ νƒœν‰μ–‘ μ§€μ—­ 10,000개 μ΄μƒμ˜ μˆ™μ†Œμ—μ„œ μ—°κ°„ 540만 건 μ΄μƒμ˜ μ˜ˆμ•½μ„ μ²˜λ¦¬ν•©λ‹ˆλ‹€.

πŸ€– μ˜€ν”ˆμ†ŒμŠ€ ML 연ꡬ

μˆ™λ°•μ—… μ½˜ν…μΈ μ˜ μžλ™ ꡬ성 및 λΆ„λ₯˜λ₯Ό μœ„ν•œ λ©€ν‹°λͺ¨λ‹¬ AI μ‹œμŠ€ν…œμ„ κ°œλ°œν•˜κ³  μžˆμŠ΅λ‹ˆλ‹€.

λ©€ν‹°λͺ¨λ‹¬ 객싀 ν΄λŸ¬μŠ€ν„°λ§

λΉ„μ „ 및 μ–Έμ–΄ λͺ¨λΈμ„ ν™œμš©ν•˜μ—¬ ν˜Έν…” 객싀 이미지와 ν…μŠ€νŠΈ μ„€λͺ…을 μžλ™μœΌλ‘œ κ·Έλ£Ήν™”ν•©λ‹ˆλ‹€.

이미지 μœ μ‚¬λ„

  • EfficientNetV2-S λ°±λ³Έ 기반 Siamese λ„€νŠΈμ›Œν¬
  • 200만 개 μ΄μƒμ˜ 이미지 쌍으둜 ν•™μŠ΅
  • μŠ€νŽ™νŠΈλŸ΄ ν΄λŸ¬μŠ€ν„°λ§μ„ ν†΅ν•œ μžλ™ κ·Έλ£Ήν™”
  • ONNXλ₯Ό ν†΅ν•œ ν”„λ‘œλ•μ…˜ 배포

ν…μŠ€νŠΈ μœ μ‚¬λ„

  • ν•œκ΅­μ–΄ 객싀λͺ… 처리λ₯Ό μœ„ν•œ KLUE BERT 기반 Siamese λ„€νŠΈμ›Œν¬
  • 31,452개 ν•™μŠ΅ 쌍으둜 νŒŒμΈνŠœλ‹
  • ν•œκ΅­μ–΄-μ˜μ–΄ λ‹€κ΅­μ–΄ ν…μŠ€νŠΈ 처리

DINOv2 λ©€ν‹°ν—€λ“œ λΆ„λ₯˜

85.6% 정확도λ₯Ό λ‹¬μ„±ν•œ ν˜Έν…” 이미지 μž₯λ©΄ λΆ„λ₯˜ μ‹œμŠ€ν…œμž…λ‹ˆλ‹€.

μ•„ν‚€ν…μ²˜

  • Vision Transformer (DINOv2) λ°±λ³Έ
  • 3개 λΆ„λ₯˜ ν—€λ“œ: Scene (6 클래슀), Concept (3 클래슀), Object (12 클래슀)
  • 127개 μˆ™μ†Œμ˜ 33,762개 μ΄λ―Έμ§€λ‘œ ν•™μŠ΅

μ„±λŠ₯

  • Scene: 85.2% | Concept: 92.8% | Object: 78.9%

ν”„λ‘œλ•μ…˜ ML νŒŒμ΄ν”„λΌμΈ

  • λͺ¨λΈ μ΅œμ ν™”: PyTorch β†’ ONNX λ³€ν™˜
  • 배포: AWS SageMakerλ₯Ό ν†΅ν•œ ν™•μž₯ κ°€λŠ₯ν•œ μΆ”λ‘ 
  • ν”„λ ˆμž„μ›Œν¬: PyTorch, Transformers (Hugging Face), timm, scikit-learn

πŸ”¬ 연ꡬ 관심사

  • μ‹œκ°μ  μœ μ‚¬λ„λ₯Ό μœ„ν•œ λŒ€μ‘° ν•™μŠ΅
  • 자기 지도 ν•™μŠ΅ λΉ„μ „ λͺ¨λΈ (DINO, DINOv2)
  • λ‹€κ΅­μ–΄ NLP (ν•œκ΅­μ–΄, μ˜μ–΄, 일본어)
  • ν”„λ‘œλ•μ…˜ μ‹œμŠ€ν…œμ„ μœ„ν•œ 효율적인 λͺ¨λΈ 배포

🌏 더 μ•Œμ•„λ³΄κΈ°

ν™ˆνŽ˜μ΄μ§€: www.onda.me

datasets 0

None public yet