PRISM

Inspiration

PRISM was built to be a "universal connection engine." The goal was to move beyond surface-level labels and use machine learning to capture the deep, intuitive patterns of human personality and behavior.

What it does

It creates a unique user "fingerprint" by analyzing interactions across 10 questions. Using a pairwise relational model, it identifies subtle traits to match like-minded individuals through high-dimensional similarity mapping.

How we built it

Developed using GAMMA, the technical pipeline includes:

• Encoding: Questions and answers are converted into 384-dimensional vectors.
• Interaction: A 10x10 matrix is generated via dot-product similarity between all Q&A pairs.
• Calibration: The signal is refined using learned scaling, bias, and ReLU activation.
• Aggregation: Data is compressed into a 64-dimensional vector for fast, scalable matching.

Challenges we ran into

• Signal Clarity: Overcoming "blurred" raw data by implementing a specific calibration step to sharpen relational signals.
• Efficiency: Optimizing the model to be significantly smaller and faster than standard transformers while retaining complex recognition capabilities.

Accomplishments that we're proud of

• Scalability: Delivering nearly instantaneous similarity results.
• Real-World Impact: Successfully deploying the engine across three distinct platforms: Hearth (support groups), Bridge (political dialogue), and Gravitas (verified local matches).

What we learned

• Relational Value: Pairwise interactions (how an answer relates to every question) provide a much richer profile than evaluating answers in isolation.
• Refinement: Learned weights are crucial for converting raw mathematical similarities into meaningful human insights.

What's next for PRISM

• Hearth: Scaling emotional-position matching for recovery and peer support.
• Bridge: Enhancing tools to humanize political discourse through shared values.
• Gravitas: Integrating AI to suggest optimal times and activities for local, verified hangouts.

Built With

• flask
• npm
• python
• pytorch