DocVision – Next-Generation Document AI
with Multimodal Data
August 26, 2025
.jpg.webp)
At Upstage, we call our colleagues “Stars.” Today, more than 140 Stars are working together on a journey to build AI that changes the world.
Mini Starview is a special interview series where Stars from different roles share their work, experiences, and growth stories in their own words.
If you're curious about how we work at Upstage, our culture, and the real growth journeys our Stars have experienced, we invite you to explore
this Mini Starview.
We hope this content provides helpful insight for future Stars considering joining Upstage and serves as a meaningful guide as you shape your
career path.
In this edition of Mini Starview, we introduce the DocVision team, which is leading the next generation of Document AI powered by multimodal
data. Today, we meet Eric Sangchul Jung, the team leader of DocVision, who drives the continuous advancement of Upstage’s models.
Q. Hello! Could you briefly introduce yourself?
Eric: Hello, I’m Eric from the DocVision team at Upstage. I focus on researching and developing next-generation Vision-Language Models based on multimodal data such as documents, images, and charts.
Q. What are the main responsibilities or projects you handle in your current role?
Eric: I work on the VLM team, where I design large-scale multimodal data collection and curation pipelines and conduct research on
training and optimizing document-centric large models based on these datasets. In particular, I focus on developing models that can
understand diverse visual elements within documents—such as tables, charts, and multi-page layouts—and perform integrated
reasoning with textual information. In parallel, I design evaluation methodologies and internal benchmarking tools that closely reflect
real-world usage scenarios, enabling continuous validation and improvement of model performance.
Q. What has been the most challenging or rewarding moment in your work?
Eric: Processing high-resolution document images requires addressing many constraints such as GPU memory limitations,
computational costs, and I/O bottlenecks. Designing the most efficient training pipeline was therefore a major challenge.
We went through many iterations while optimizing the entire workflow—from data loading and preprocessing to model input.
Eventually, we successfully built a pipeline capable of stably training large-scale multimodal datasets.
As a result, both training speed and resource efficiency improved significantly, which ultimately contributed to better model
performance. That experience was particularly rewarding.
Q. What is the most valuable lesson you’ve learned from collaborating with colleagues in the same role?
Eric: The biggest lesson I’ve learned from collaborating with colleagues is the importance of listening to and respecting each other's
perspectives. In research, many different approaches are proposed, and sometimes opinions conflict.
However, our team members maintain a culture where everyone listens with an open mind and works together to find the most
reasonable solution.
Through this process, individual ideas come together toward a shared goal, and I’ve experienced what true one-team collaboration
feels like.
Q. How have you grown since joining Upstage?
Eric: The biggest change for me has been my perspective on work.
Startups have a high level of energy density, where everyone must immerse themselves quickly and deliver results.
On top of that, our team aims to build world-class technology.
Because of that, results that I might have previously considered “good enough” are now viewed from a different perspective:
“What more can we try to achieve even better performance?” This mindset has helped me grow significantly as an engineer.
Q. What do you hope to achieve at Upstage in the future?
Eric: I want to build Document AI models that can be used globally, enabling both companies and individuals to maximize their
productivity through AI.
Q. How does your team collaborate and work with other teams?
Eric: Our team actively uses multiple communication channels.
For quick discussions or idea sharing, we communicate through Slack, while important information is systematically documented in
Notion and used as the basis for meetings.
This approach helps reduce unnecessary meetings while ensuring that essential discussions remain focused and efficient.
A documentation-driven workflow also improves shared understanding within the team and enables smoother collaboration with
other teams and functions.
Q. Was there anything that impressed you when you first joined Upstage?
Eric: When I first joined, the fully remote working environment felt a bit unfamiliar.
I remember being surprised when meetings ended and everyone simply clicked the “leave” button and disappeared at the same time.
However, over time, online meetings began to feel just as natural and comfortable as discussing something with colleagues sitting
nearby. Now, I actually feel that my work efficiency is even higher than when working in the same physical space.
At this point, I’m so satisfied with the remote setup that it would be hard to imagine working any other way.
Q. Do you have any practical advice for candidates interested in this role?
Eric: One important thing is to think deeply about the problems Upstage is trying to solve.
When you understand those problems in depth, it becomes clearer what kinds of technical challenges need to be addressed and how your previous research or learning experiences could contribute.
Through that process, I encourage candidates to consider what role they could play at Upstage.
Q. Is there anything you’d like to say to candidates considering applying to Upstage?
Eric: I believe Upstage is one of the most fast-paced and technically deep companies that is genuinely striving to make the world
better through technology.
The AI field evolves rapidly, and to keep up—or sometimes move ahead—everyone here invests their full capabilities.
For that reason, I believe Upstage is one of the few AI technology companies in Korea that can truly compete on the global stage.
I hope you’ll join us on this journey and help create that change together.