Seonglae Cho일단 호감으로 보이게 잘웃고 여유롭고 말을 똑바르게 끝까지 한단어한단어 입 크게 움직이며
I think 같은 버리는말 하지말기
별거 아니니 생각하면서 말하기
my best part is engineering eventually make it by any method
solid CS knowledge and Web ecosystem understanding with 3 years of professional experience leading to build web software just as UIpath does
also responsibility to commitment about the company considering priority
이름: 미루나
Good morning! How are you? Is ‘Miruna’ the correct pronunciation of your name?
My name is Seonglae Cho, but you can also call me Sung if that's easier
- Tell me about yourself and your experience in AI/ML.
Hello, I’m Seonglae Cho, and I’m currently completing an MSc in AI at UCL. Previously, I earned my Bachelor’s in Computer Science and spent three years as a Software Engineer in the autonomous driving sector, where I contributed to a 3D dataset pipeline for self-driving cars. To make more fundamental contributions in AI, I chose to study at UCL and am now doing my thesis on improving AI reasoning capabilities by leveraging reinforcement learning. I also continue to contribute to open-source projects and have published my own web services using machine learning and information retrieval techniques.
- MSc
- Bachelors
- Kakao pipeline
- 2
- open source
- Describe a real-world problem you solved using machine learning—what was your approach and outcome?
MBTI GPT AI solution
To apply machine learning, data is essential, but when tackling new problems, relevant data often doesn't exist. Despite this challenge, I have successfully applied machine learning in several research projects, most notably in my NAACL paper RTSum. In AI-based summarization, the main goal is to extract and preserve core information accurately. However, AI models often hallucinate — they may fabricate information or produce inaccurate content. To address these issues, I decomposed sentences into smaller phrase units and then reconstructed them back into full sentences. This approach helped mitigate both information loss and hallucination problems.
This work was conducted in an unsupervised manner by training on a self-generated dataset. In my recent, yet-unpublished paper, I also guided an LLM to create a self-generated dataset and extracted interpretable token distributions. Through this approach, I demonstrated how unsupervised methods can address the lack of labeled data in niche markets with high sample efficiency.
I believe this experience would directly help me solve problems in environments like UiPath, where data scarcity can be a critical challenge.
- How do you handle feature engineering and selection for a new dataset?
First, visualization. For each dataset, I create an appropriate visualization and build my own mapping table to quickly understand the dataset’s characteristics.