Core AI Research & Technologies
AI research that understands content context and specialized knowledge
Research direction
We research AI that analyzes content, connects it to knowledge, and supports work. Our culture research is underway, with joint research planned to extend the foundation into healthcare.
National R&D projects and large-scale AI datasets have helped us validate capabilities and extend our research into new fields.
Research into user-specific machine translation, source-text correction, and translation document management built the foundation of TWIGFARM’s language AI.
Scope widened to multimodal machine translation that reads video, audio, text, and cultural context together, alongside broadcast subtitle production, editing, and content analysis.
Work on AI Hub datasets accumulated the ability to design, refine, and verify text, speech, image, video, and specialized-domain data.
Semantic search, multimodal metadata generation, media sales kits, and AI-driven content asset management are being applied in LETR WORKS.
Starting with culture and healthcare: LLMs that understand expert knowledge and the context of a working environment, and that support decisions with traceable evidence.
We analyze video, audio, subtitles, scripts, and images to structure people, actions, dialogue, and mood. The results support semantic search, summarization, scene recommendations, subtitling, dubbing, and content reuse.
We research language AI that accounts for genre, character relationships, tone, and cultural context. Video scenes, speech, facial expressions, and existing translation assets help preserve meaning and emotion.
We design data for model training and evaluation. Automated refinement, expert review, quantitative evaluation, and model-based checks help establish data quality.
We structure documents, databases, and business rules with ontologies and retrieval so LLMs can answer with checkable evidence. This research extends to domain agents that connect retrieval, analysis, verification, reporting, and external tools.
We research collaboration systems that combine AI processing with expert judgement. Corrections and approvals inform customer terminology, evaluation criteria, and workflows to improve quality and productivity.
Capabilities developed through R&D and data projects support current services and further research.
| Validated result | Capability developed | Research direction |
|---|---|---|
| Machine translation and source-correction R&D | Language understanding, custom translation | Culture-aware LLM |
| Multimodal machine translation R&D | Video, audio, and text fusion | Cultural reception model |
| Broadcast subtitle R&D | Media processing and workflow | Multimodal content intelligence |
| K-Stock dataset | Cultural image and text data | Culture-specific multimodal models |
| Law, finance, and defense datasets | Specialized-domain data modeling | Domain-specific LLM |
| AI Hub dataset construction | Large-scale data quality management | Training and evaluation data for culture and healthcare |
| LETR WORKS commercialization | Turning research into product | Industry-specific AI platforms |
Both fields share multimodal understanding, domain knowledge, reliability evaluation, and expert review. Culture research examines content interpretation and reception; planned healthcare research focuses on professional decision support.
We study how countries, languages, generations, and cultures affect the interpretation and reception of the same content. K-Stock data carrying Korean cultural, everyday, and geographic context and metadata describing background and intent provide a research foundation.
We plan joint research on healthcare LLMs with hospitals, universities, and specialist institutions. TWIGFARM would contribute data pipelines, RAG, evaluation, and operations, while partners would define clinical problems and provide medical data and expert validation.
The healthcare work is a planned joint research direction, not a commercially released product. It targets decision support that provides evidence and information to clinicians and care professionals, without replacing diagnosis or clinical judgement.
Research reaches services through data development and validation. Feedback from real use guides further improvements.
New models, data structures, and evaluation methods.
Source data from the field, designed and refined into a structure that can be trained and evaluated on.
Performance and stability verified through automated evaluation, expert review, government projects, and joint research.
Validated technology applied in LETR WORKS or in a customer-specific AI system.
Corrections and evaluation results from live use fed back into customer-specific data, rules, and models.
Results from data projects and applications in commercial services.
Culture-grounded multimodal AI data
Image and text data carrying Korean cultural and everyday context, designed and then verified with model-based quality checks.
Multimodal data and clean media
K-Stock content data plus de-identified, clean-copy broadcast video widened the base for multimodal data use.
Research reaching the service
Semantic search over the meaning of video, audio, and text, applied as LETR WORKS’s content asset management feature.
From language AI to media intelligence
How translation and localization technology grew into multimodal content asset management and a global monetization platform.
Cloud-based AI infrastructure processes large media and specialized datasets. Modular collection, preprocessing, inference, quality checks, and deployment help bring research into industry use.
GPU-based training and inference use distributed processing designed to scale horizontally with data volume.
We manage model and data versions together and track results through performance monitoring and data drift detection.
We apply access controls, audit logs, and isolated customer workspaces, alongside incident response and backup/recovery procedures. Specific availability levels are agreed in the service contract.
We seek joint research with institutions and companies that bring expertise in culture, content, medicine, and care. Together, we can define AI datasets, evaluation metrics, and validation procedures.
