Inspiration
Inspiration comes from the movie "Her" and its futuristic concept of having digital companionship through AI and it's impact on an individuals mental health
What it does
It's a simple ai agent that has a proprietary NLP model as a tool that analyzes user emotional state and helps the AI model understand the users mental condition and help accordingly. The model can classify seven emotional state classes (Normal, Depression, Suicidal, Anxiety, Stress, Bi-Polar, Personality Disorder) and give a percentage score range among all the classes that add upto 100 total. Each probabilty helps agents decipher user emotion for that current text input.
How we built it
For the NLP model we used a hybrid architecture of Vector Embeddings to generate semantics and train a XGBoost classifier on those embeddings to predict all the target classes. Dataset used to train this model is- "https://www.kaggle.com/datasets/suchintikasarkar/sentiment-analysis-for-mental-health", the embedding model used is- "all-MiniLM-L6-v2"
Challenges we ran into
The dataset was incomplete and imbalanced, so we used distilledbert and gpt-oss 20B models to generate synthethic and diversify the data to fill in the blanks and make the data somehwhat usable. But the fact remains that the model will still produce wrong predictions due to the lack of proper data because the dataset used is relatively small and still needs more diverse contributions for each individual classes for more diversity in human test expressions.
Accomplishments that we're proud of
the model produced is extremely fast and moderately accurate for small to medium text paragraphs (200-300 characters) and compatiblity as a agentic tool is working very smoothly.
What we learned
We learned the critical importance of data preprocessing and balancing when dealing with sensitive, small datasets. Utilizing DistilBERT and open-source GPT models for synthetic data generation taught us how to safely augment text data. Finally, we learned how to effectively pair semantic vector embeddings with an XGBoost classifier to create a lightweight, fast, and highly responsive tool for AI agents.
What's next for Moodly
Next, we want to expand our dataset with diverse, real-world examples to improve classification accuracy and reduce errors. Ultimately, we aim to introduce multi-modal capabilities like voice and facial analysis for a more holistic digital companion.
Log in or sign up for Devpost to join the conversation.