Inspiration
Emotions can be overwhelming, complex, and incredibly difficult to manage alone. The struggle often intensifies when seeking guidance from well-meaning but unqualified friends or colleagues. This can lead to biased, weak, or even counterproductive assessments of one's emotional state, deepening the dilemma rather than resolving it. While professional help is profoundly effective, the prospect of sharing deeply personal emotions with a stranger can feel like a daunting, high-stakes step. This hesitation often prevents individuals from seeking the support they need. As a result, many who could significantly benefit from professional intervention, may choose to keep to themselves, creating a self-imposed barrier, perpetuating a cycle of emotional distress. Our guiding mission: To reduce the self-imposed barrier to seeking professional help for emotional health, while providing easily-accessible, transparent and verifiable, emotionally-uplifting experiences for short-term relief.
What it does
Users freely express their feelings within their private, secure VR environment. This space ensures complete confidentiality and comfort, fostering open communication. Our application leverages advanced OpenAI models to accurately transcribe the user's spoken words into text in real-time An LLM performs an initial sentiment analysis, identifying and classifying the user's emotional state (e.g., "Frustrated," "Anxious," "Sad") An LLM first redacts all Personally Identifiable Information (PII) from the user's conversation, ensuring complete privacy and data protection. The user is presented with the redacted text for review and explicit approval. Only after user consent, is the anonymized conversation made viewable to registered experts. Qualified experts review and comment on the AI's sentiment analysis and suggested gamified experiences. Their professional feedback and validation are then made transparently visible to the user. Our platform is supposed to serve as a bridge to professional, clinically approved emotional therapy, not a replacement. However, user may choose to never share the history of interaction on the platform despite PII redaction. In that case, the number of times user can avail AI sentiment analysis and subsequent soothing experiences, is limited The information redaction and/or user review before data sharing, might change the context of conversation too much for correct audit by experts. In that case too, we want to have limits on how many times user can get analysis from our platform for a continuing emotional state (“Sad” emotion detected more than 3 times, for example). This reduces the chances of expert’s advice being taken out of context and expert being responsible for a mis-diagnosis due to lack of context The WHO estimates anxiety & depression cost the global economy $1 Trillion in lost productivity. Deloitte's research indicates that for every £1 invested in workplace mental health initiatives, businesses see an average return of £4.70, which translates to a 5:1 ROI by reducing absenteeism and employee turnover. We sell this ROI to corporate entities The requirement of gamified experiences on the platform is fulfilled by game developers willing to sell their games, either as one-time sale or rentals. Special sales plan for bulk purchases by corporate clients can be set up. Our platform’s game marketplace facilitates this purchase and takes a cut. Emotional/mental health experts reviewing AI analyses on the platform are rewarded through exposure to potential clients. When number of AI sentiment analyses are limited for users, they can choose to be handed off to one of the experts that reviewed their case. This hand-off can bring our platform a referral reward from the expert.
How we built it
The application is supplied through Unity-based VR interface suitable for VR headsets. The application content is overlayed onto user’s real-world POV and user interaction is maintained parallel to user’s current tasks Leverages state-of-the-art OpenAI models for highly accurate real-time transcription, sentiment analysis of user input and PII redaction Offers one of several available gamified activities available to our Unity application’s Python (Cuify by TUM Think Tank) backend, designed to soothe emotional states. LLMs are used to choose the game based on mood detected during current sentiment analysis Streamlit (Python) interface to allow user to choose the conversation and associated AI analysis they want to make public for review, through explicit consent mechanisms Streamlit (Python) interface for certified professionals to validate AI assessments, and provide feedback, on data made public by user
Challenges we ran into
Getting to know Unity framework as a beginner Navigating legal and ethical issues related to pitch
Accomplishments that we're proud of
A well thought of UX and business revenue model and working prototype in limited time
What we learned
Unity framework Nuances of using VR experiences of technology in general, in health care
What's next for EmoBuddy
The application along with VR hardware, can be used to create a focus zone limiting distractions, for users suufering from procrastination, doom-scrolling, etc. Users may choose a particular screen with a particular program (a work presentation for example) that they want to focus on. The screen can be focused more in VR environment while surroundings are blurred (not while moving). Any action by user to step out of focus zone before specified time limit can be interrupted by gamified exercises to calm down urge for dopamine gains via distractions like social media In current setup OpenAI models hosted on cloud are needed for speech-to-text conversion, text-to-speech conversion, sentiment analysis, game allotment and PII redaction, for smooth and fast responses for demonstration purposes. But in some cases, sentiment analysis could be required as a once-a-day journaling activity only. In that case slow, but accurate local AI models can be enough. We evaluated WhisperX model for speech-to-text task and reported transcription of 120-second-long audio in under 40 seconds, without GPU in consumer-grade laptop. Piper-TTS model provided various languages and accents, for local non-GPU text-to-speech tasks, with almost instantaneous results. GPT-OSS-120B open-source LLM hosted on Ollama locally could carry out sentiment analysis and PII redaction in around 15 minutes. These options can improve the control over sensitive data and also drive down costs for using the platform, in certain use-cases.
Log in or sign up for Devpost to join the conversation.