Inspiration
We are currently facing a silent epidemic on college campuses: loneliness. Recent data shows that a staggering 60% of college students report feeling "very lonely," and 44% experience symptoms of depression. While there are countless social apps on the market, platforms like Bumble BFF or Friended rely on static questionnaires or surface-level interests. Shared hobbies do not necessarily equal emotional resonance.
We asked ourselves: What if we could use AI to understand a student's evolving emotional state and match them with peers who truly resonate with them on a deeper, psychological level? This inspired NeuroTwin, a system that moves beyond static profiles to create a living, breathing Emotional Digital Twin for every student, fostering authentic connections when they need them most.
What it does
NeuroTwin is an emotionally intelligent digital twin system designed to build meaningful peer connections.
Instead of a one-time setup, NeuroTwin learns about a user continuously. Through private daily journaling, mood check-ins (energy, stress, and social battery sliders), and activity logs, the platform constructs an emotional model of the student.
When a student is looking for connection, NeuroTwin's AI computes an Emotional Compatibility Score against other students on campus. It filters out the noise, recommending only peers who surpass a 50% compatibility threshold. To eliminate the anxiety of the "first message," our system also generates a context-aware, personalized icebreaker for every match, giving students a safe and natural way to start a conversation.
How we built it
We built NeuroTwin with a modern, scalable, and AI-native stack, focusing heavily on semantic understanding rather than basic keyword matching.
- Frontend: A responsive, accessible single-page application built with Next.js 16, TypeScript, and Tailwind CSS 4.
- Backend: A highly performant, type-safe REST API powered by FastAPI and Python 3.12.
- Database & AI Engine: We leveraged Snowflake for our cloud-native database and Snowflake Cortex AI to generate vector embeddings (
EMBED_TEXT) and extract emotional sentiment (COMPLETE,SENTIMENT). - Orchestration: We used LangChain Runnables to manage our three core AI pipelines:
- TwinBuilderChain: Ingests journals and updates the twin snapshot.
- MatchRetrievalChain: Performs vector searches for candidates.
- ExplanationChain: Generates human-readable explanations and icebreakers.
To accurately match students, we developed a multi-dimensional Compatibility Scoring Model. We implemented this using the following weighted formula to ensure a holistic match:
Compatibility Score (%) = 100 * clamp( (0.50 * Cosine_Similarity(A_embed, B_embed)) + (0.20 * Emotion_Similarity) + (0.15 * Activity_Overlap) + (0.10 * Values_Alignment) + (0.05 * Schedule_Overlap) )
Challenges we ran into
Privacy vs. Personalization: Our biggest hurdle was figuring out how to train a highly accurate emotional matching system without exposing highly sensitive, personal journal entries. We solved this by implementing a Zero Raw Sharing architecture. Journal text is Fernet-encrypted at rest. The LangChain pipeline extracts only aggregated emotional patterns and vector embeddings. Raw text is never shared with the matching algorithm or other students.
Algorithmic Weighting: Initially, our compatibility scores were heavily skewed by shared activities rather than actual emotional states. Fine-tuning the cosine similarity weights of the Snowflake Cortex embeddings against Jaccard similarities for activities and values took significant iteration to get right.
Accomplishments that we're proud of
- Pioneering Emotional Matching: We successfully built the first platform that matches users based on evolving emotional patterns rather than static traits.
- Privacy by Design: We implemented strict data protection, including explicit consent management, GDPR-compliant data deletion endpoints, and encrypted storage, proving that AI applications in the mental health space can be both powerful and secure.
- Accessibility: We didn't treat accessibility as an afterthought. NeuroTwin features full screen-reader support, semantic HTML, comprehensive keyboard navigation, and reduced-motion states for neurodivergent users.
What we learned
We learned that Large Language Model embeddings are incredibly powerful for semantic emotional understanding. Traditional matching algorithms look for identical keywords; our vector-based approach understands that a student feeling "overwhelmed by finals" and a student feeling "anxious about my workload" share an emotional resonance, even if they use different words.
We also learned that in the mental health technology space, trust is your actual product. Building a robust, transparent privacy architecture was just as important as building the AI model itself.
What's next for NeuroTwin
This is just the beginning of our vision to eradicate campus loneliness. Our immediate roadmap includes:
- Institutional Dashboards (v2.0): Creating anonymized, aggregate dashboards for university counselors to identify macro-trends in student mental health without compromising individual privacy.
- Mobile Expansion: Developing a React Native mobile application to allow for offline journaling and real-time push notifications for new matches.
- Group Dynamics: Expanding our matching algorithm to support "interest clusters" and emotionally compatible study groups, bringing NeuroTwin's power to collaborative campus environments.
Log in or sign up for Devpost to join the conversation.