Built an LLM that mimics love language in Korean, trained on 270B tokens from 100K+ real Korean couple messenger chats, to research AI's ability to mimic emotions in conversations. Team size: 4 ●​ Accelerated context vector computation for 3M messages by distributing workloads across 4 GPUs, achieving up to 4x speedup. ●​ Reduced the dimensionality of 1M message vectors and visualized them to identify unique trends in message distributions. ●​ Co-developed a feature to predict the time taken to get a reply to a message by fine-tuning the model with 1B tokens, improving prediction accuracy from 12.5% up to 40%. ●​ Built a distributed data processing pipeline, expanding the dataset from 10B to 270B tokens, reducing overfitting issues. ●​ Stack: Python, PyTorch, CUDA, Linux, Docker, Kubernetes, numpy, pandas

더보기

Co-developed a multilingual event planning platform with a chatbot, enabling users to plan events via messengers. Team size: 5 ●​ Enhanced security for the admin panel by implementing an authentication, significantly reducing unauthorized access attempts. ●​ Decreased database query latency by up to 25% by populating an in-memory H2 database, improving performance. ●​ Developed CRUD functions, enabling users to interact with the database in the chatbot, streamlining the user experience. ●​ Implemented unit testing for the 5 main functionalities using the Mockito framework, enhancing test coverage and robustness. ●​ Stack: Java, Spring Boot, PostgreSQL, Mockito, Linux

더보기

OnSquare - Company behind Soundcorset, used by 30M+ users on Google Play. Team size: 7 ●​ Reduced editing time in a hex editor by implementing copy, cut, and paste features achieving user productivity boost up to 50%. ●​ Improved cross-platform graph compatibility by adjusting labels and figures for screen sizes, boosting user satisfaction. ●​ Reduced user clicks from 3 to 1 by automating with customizable shortcut buttons, increasing task completion speed up to 30%. ●​ Stack: TypeScript, Redom, d3.js, Linux

더보기