Capstone Project: AI Emotional Support Service for the Socially Vulnerable

💙 Daily Mood: AI Emotional Support Service for the Socially Vulnerable

Team Project | Sep 2024 – Dec 2024

Proposed and built “DailyMood”, an AI-based emotional diary analysis service designed to support socially vulnerable groups (youth, elderly, and isolated individuals).
Implemented a KoBERT-based emotion classification model with SentencePiece tokenizer, AdamW optimizer, and CrossEntropyLoss, fine-tuned on AI Hub dialogue datasets (6-class emotion taxonomy: Happiness, Sadness, Anger, Fear, Surprise, Disgust).
Integrated a KoBART-based summarization model, fine-tuned on Korean summarization datasets, to generate concise summaries of user diaries.
Applied hyperparameter optimization (batch size 64, learning rate 3e-5, gradient clipping, linear scheduler) and experimented with Mixed Precision Training for efficient GPU utilization.
Designed the system pipeline: diary input → backend transfer → NLP-based summarization & sentiment classification → frontend visualization.
Developed a Flutter + Spring Boot prototype app, displaying diary entries, emotion statistics (weekly/monthly), and personalized summaries.
Built data pipelines for preprocessing (tokenization, dataset augmentation, and JSON → CSV conversion) to handle over 180K+ Korean text samples.
Conducted performance comparison between different training setups, analyzing model accuracy vs. real classification quality, and selected the best performing configuration.
Final prototype showcased via interactive demo video and team presentation, highlighting both technical contributions and societal impact.