Onda Launches Hospitality-Dedicated Multimodal AI on Hugging Face

ONDA, an innovative company in the hospitality AI sector, has garnered attention by fully releasing its self-developed multimodal AI model on Hugging Face, a global AI platform. This model is a hybrid AI that combines a self-built room image dataset of over 8,000 images with Korean-specific natural language processing technology, achieving an astonishing accuracy of 96.5% in solving the chronic data inconsistency problem of accommodation distribution platforms. Currently, accommodation distribution platforms, including OTAs (Online Travel Agencies), collect room data for the same hotel from multiple suppliers, but data inconsistency issues arise because each supplier uses different room naming conventions. For example, the same room might be listed by supplier A as ‘Deluxe Room – 1 King Bed (West Tower)’ and by supplier B as ‘King Room – Club Access, West Tower’. These inconsistencies lead to duplicate inventory creation, image mismatches, and customer confusion, acting as a persistent obstacle to platform operations. ONDA AI Lab built a unique hybrid AI architecture to solve this complex problem. A Vision Transformer (ViT)-based image classification model precisely recognizes room components such as bedrooms, living rooms, and bathrooms, while a Korean-specific language model (KLUE BERT) analyzes subtle semantic differences in specialized hospitality domain terms like ‘Deluxe,’ ‘Ocean View,’ and ‘Early Check-in.’ Through an ensemble technique that weighted-fuses the predictions of both models, it maintained accuracy by identifying even subtle differences that are hard to discern with single pieces of information, such as rooms with the same bed image but different views. Instead of blindly trusting AI predictions, ONDA maximized practicality by introducing a reliability-based hybrid workflow. Each prediction result generated by the model is assigned a confidence score, and if a score falls below a set threshold, a hospitality domain expert conducts the final review through an AI-expert collaboration system, ensuring stability and reliability in real-world environments. This release on Hugging Face goes beyond a mere display of technology. It transparently demonstrates that ONDA independently carried out the entire process, from dataset construction to model training, evaluation, and deployment, firmly showcasing the authenticity of its AI technological capabilities. The vast real transaction data accumulated since its establishment in 2016, connecting 37,000 domestic accommodation providers and 71 distribution channels, served as a solid foundation for this AI model development. This technology is expected to expand beyond automatic room attribute tagging, price optimization, and inventory forecasting to more diverse AI-based automation solutions such as automatic recognition of room amenities, view type classification, and image quality assessment. Oh Hyun-seok, CEO of ONDA, emphasized, "True AI competitiveness comes not from the model itself, but from high-quality domain data to train it and the ability to apply it in practice," articulating ONDA's unique position to implement 'AI that actually works,' based on its deep understanding of the hospitality industry and 8 years of accumulated data.