ReCall.AI
Princeton Hackathon • Winner in 2 Categories
The Problem
Alzheimer's patients often struggle to recognize familiar faces, causing distress for both patients and their loved ones. Existing memory aids were static and couldn't adapt to real-world scenarios where patients encounter people in various contexts and lighting conditions.
Challenge
In just 48 hours, we needed to create a functional application that could perform real-time facial recognition, match faces to stored memories with high accuracy, and present information in a way that would be helpful rather than overwhelming for Alzheimer's patients.
Solution
We created a compassionate AI application that uses computer vision to help Alzheimer's patients recognize loved ones. The system captures faces in real-time, converts them to vector embeddings using DeepFace, searches a PineCone vector database for matches, and displays relevant information about recognized individuals in a clear, non-overwhelming format.
Process
- Rapid Prototyping: Researched facial recognition libraries and vector database options. Set up basic DeepFace integration and tested accuracy with sample images.
- Core Development: Built multi-threaded facial recognition pipeline with real-time processing capabilities. Implemented PineCone vector database for efficient similarity search.
- Integration & Testing: Integrated ML backend with user interface. Tested with various lighting conditions and face angles to ensure robust recognition.
- Demo Preparation: Refined user experience and prepared compelling demo showcasing real-world use cases for judges and audience.
Challenges Overcome
- Real-time Performance: Implemented multi-threading to handle facial recognition processing without blocking the user interface, ensuring smooth real-time operation
- Recognition Accuracy: Used FlatL2 indexing and optimized embedding parameters to achieve high accuracy even with varying lighting and angles
- User Experience for Patients: Designed simple, clear interface with large text and familiar visual cues that wouldn't confuse or overwhelm users
Impact & Results
- Hackathon Outcome: Concept → Winner in 2 tracks (Top recognition from judges working at Meta, Amazon, Princeton, etc.)
- Recognition Speed: N/A → Real-time processing (Instantaneous face matching)
- Database Efficiency: Traditional storage → Vector similarity search (Optimized for facial recognition)
Key Achievements
- Won in 2 categories at Princeton University hackathon with innovative Alzheimer's care solution
- Built multi-threaded facial recognition system with real-time processing capabilities
- Implemented efficient vector similarity search using DeepFace and PineCone database
- Created patient-centered interface designed specifically for Alzheimer's care needs
Tech Stack
DeepFace, PineCone Vector Database, Facial Recognition, Vector Embeddings, Multi-threading, Computer Vision, FlatL2 Indexing, Real-time Processing