I am AI researcher and engineer based in Seoul, South Korea. I hold a PhD degree in computer science and have expertise in machine learning, deep learning, and AI. I looking forward to seeing AI transformation in every corner of our lives that we have been witnessing for the past years. I am always looking for real-world challenges where AI can empower and help people.
In my free time, I like writing scripts to automate my workflows, some of which you can find in my dotfiles or agent-slash-cmd Git repository. I enjoy rock climbing, specifically indoor bouldering. I am deeply concerned about the climate crisis, so I became vegan because it's the least I can do.
I joined Mind AI as a senior AI Scientist. I liked the experience at Mind AI because the team was very agile and pivoted a lot following the latest trends of LLMs and agents. I also met a great team lead who had deep understanding of AI. Mind AI's focus was to augment LLM's capabilities either through logical validation, prmopt engineering, context engineering, or harness engineering. Thanks to the agentic coding wave, we were able to build many prototypes and proof-of-concept products.
What I am most proud of, as a senior AI scientist, is that I voiced my ideas and opinions, and contributed to the direction of the company's vision and products.
Highlights
I led the ML research and engineering team at NoriSpace. NoriSpace offers ML-powered solutions to automate business tasks. Business data allows access to massive and diverse data and involves many modes, such as texts, images, a lot of documents (images with texts), event logs, etc. It is a fantastic opportunity to study, learn, build, and deploy many types of models and to manage their cycles.
Highlights
I did my PhD as a member of the advanced microscopy team in LOB. Our team dealt with large volume fluorescence microscopy images and develops non-linear microscopy. I researched and utilized self-supervised learning approach with deep neural networks to develop a generic filter for nuclei and cells, that covers various microscopy images. My advisors were Anatole Chessel and Emmanuel Beaurepaire.
It was this moment when machine learning and entrepreneurship were engraved in my brain. My course of Master's degree study was a bit longer than usual and quite jumpy. To be precise, it took me 9 months in KHU, South Korean and 20 months in École Polytechnique, France. I started off my Master's degree studying quantum engineering for future display, but came to know something called computer vision using machine learning and deep learning and it changed everything. To be honest, I cannot say I've learned a lot about entrepreneurship from classes because I didn't take any during this period. But hey, I landed an internship at an AI startup. All of my courses were about machine learning, deep learning, and computer science which I had to catch up. I would say this much that persuading my advisor professor was not easy, and I am still proud of my decision.
I worked at a start-up called PandaScore as an intern in the Data Science team. Here I met real data scientists. Working in a start-up was absolutely a fascinating experience in the middle of my study.
I came to France and first encountered machine learning and computer vision. Segmenting bio-medical fluorescence images was a cool thing. This opportunity was a turning point of my career towards computer science and machine learning.
I wanted to make a next generation display and decided to do Master's degree. In Quantum Information Display unit, I synthesized CVD (chemical vapor decomposition) graphene and put it on anodic aluminum oxide, which I grew in a custom built chemical chamber. The device showed some photo sensitivity.
I majored in what's called Information Display, an interdisciplinary engineering subject, covering electronics, optics, chemistry, and even a bit of quantum theory. You might be wondering what took me so long for Bachelor's degree. I did the military duty for 21 months. Yeah, it's still a thing in South Korea, if you didn't know...
I don't have a crazy number of agent skills for coding harnesses. But I have some that I have found useful. I tend to make custom skills on the fly for each project for certain tasks which I feel confident enough to delegate to agents. I have a couple of common skills that I use across projects, and you can find them in this repository.
I am a believer who thinks that reading is the superpower and the greatest value what my long study granted me. When people say "Oh, you have a PhD. You must be very smart". No, I wouldn't say I am smart. I just have perseverance to put time and energy on subjects that most people can't justify pouring their own time and energy. Yes, scientific papers are written in a way that sometimes they sound enigmatic. It's usually because you don't have the same context that the authors had when they wrote them. It feels rewarding when you dug up the right references and finally understood what they meant. It accumulates over time for a certain topic. Next time you pick up a paper with the same topic, you are better prepared.
This is why I track all my paper reading. I have been using Zotero, paper reference management software. Zotero is such an amazing software, and I recommended it to many people. I didn't read them all in the list from cover to cover, to be frank. I make entries for those I read their abstract. You can find my paper reading list following the link.
I became a full-time Linux user when I got into serious programming. I found that the development experience on Linux was much painless than on Windows, and in fact it was much easier to use Linux than I had anticipated. Linux gives me much more control on my system and hardware as well. I like the Unix philosophy, and I love and support open source software. My favorite distro is Arch Linux. I use a tiling window manager (bspwm), neovim, lf, and others. Also, it happens that I enjoy writing small scripts to automate things and configuring software to my taste. This project basically contains all my configuration guides and files as well as scripts to make my current setup.
bioimageloader is a python library to make it easy to load bioimage datasets for machine learning and deep learning. Bioimages come in numerous and inhomogeneous forms. bioimageloader attempts to wrap them in unified interfaces, so that you can easily concatenate, perform image augmentation, and batch-load them.
While supervised deep neural networks have become the dominant method for image analysis tasks in bioimages, truly versatile methods are not available yet because of the diversity of modalities and conditions and the cost of retraining. In practice, day-to-day biological image analysis still largely relies on ad hoc workflows often using classical linear filters. We propose NU-Net, a convolutional neural network filter selectively enhancing cells and nuclei, as a drop-in replacement of chains of classical linear filters in bioimage analysis pipelines. Using a style transfer architecture, a novel perceptual loss implicitly learns a soft separation of background and foreground. We used self-supervised training using 25 datasets covering diverse modalities of nuclear and cellular images. We show its ability to selectively improve contrast, remove background and enhance objects across a wide range of datasets and workflow while keeping image content. The pre-trained models are light and practical, and published as free and open-source software for the community. NU-Net is also available as a plugin for Napari.
S. Lim, E. Beaurepaire, and A. Chessel, "NU-Net: a self-supervised smart filter for enhancing blobs in bioimages", BIC workshop, ICCV 2023. link (Also, check out my talk for the same work in I2K conference in 2022: link)
S. Lim, X. Zhang, E. Beaurepaire, and A. Chessel, “BioImageLoader: Easy Handling of Bioimage Datasets for Machine Learning,” arXiv.org, Mar. 02, 2023. link
[Dissertation] S. Lim, E. Beaurepaire, and A. Chessel, "Versatile machine learning for neurodevelopmental imaging", Institut polytechnique de Paris, École doctorale de l'Institut polytechnique de Paris, Oct. 2023. link