Hwalsuk Lee, also known as Stan Lee, is Upstage’s co-founder and chief technology officer. His work connects computer vision, document processing and language models, following earlier leadership of Naver Clova’s Visual AI team.[1][2][6]
Start with CRAFT text detector →
Career
Lee earned a doctorate at KAIST with work on video codecs. In his university-hosted career account, he describes moving from CCTV technology into AI, then working at NCsoft before leading OCR and broader computer-vision work at Naver. He later joined Kim in building Upstage. The account emphasizes that deployed services require data collection and evaluation methods matched to real user requirements.[3]
Contributions
Lee co-authored CRAFT, a scene-text detector. Instead of representing every word as a rigid box during learning, it predicts character regions and the affinities connecting them. Those predictions can then be grouped into text regions, including curved or irregular text. The original implementation exposes pretrained models and post-processing code for examining the method.[4][5]
Ideas
In his original Hyperight interview, Lee explains why document AI needs more than recognizing characters. Tables and page structure must survive conversion into material that a language model can retrieve and use. He connects OCR, layout analysis, embeddings and retrieval with enterprise use cases, while emphasizing the quality of the underlying data.[6]
Tools, writing & talks
Tools
Start here
CRAFT text detector
github.comThe official character-region and affinity-based scene-text detector implementation.
Begin with the example detections and the character-affinity explanation. Python and computer-vision knowledge help; the repository uses older dependencies, so inspect setup requirements before running it.
Writing, research & talks
- Document AI and enterprise RAGhyperight.com
Lee’s explanation of preserving document structure when connecting enterprise data to language models.
Interviewee[6]