Back to Projects

Blindlove

Anonymous dating application with AI-powered intelligent matching using vector embeddings and semantic user profiling.

View Project
Blindlove

Project Overview

Innovative dating platform featuring an advanced anonymous matching algorithm powered by AI vector embeddings and semantic analysis. Uses Redis Vector Search with Hugging Face transformers for intelligent user profile tokenization and similarity matching. Built with real-time communication capabilities, multi-language support, and scalable architecture using Node.js clustering and Redis Stack for high-performance operations.

Role

Lead Backend Developer and AI Architect

Technologies Used

NestJSSocket.ioMongoDBSSERedis StackRedis Vector SearchHugging Face TransformersAWSJWTReact.jsPython

Tasks

  • Implemented advanced matching algorithm with pooling based on general filtration using weighted preferences
  • Applied cosine similarity and bitmap interests for match scoring
  • Integrated Redis Vector Search (RediSearch) for semantic user profile matching
  • Implemented user profile tokenization using Hugging Face sentence transformers
  • Built vector embedding pipeline converting user profiles, interests, and bios into searchable embeddings
  • Created shared/utils services module for code reusability across the application
  • Implemented event listening for service/module decoupling and maintainability
  • Built comprehensive multi-language module with custom interceptor, decorator, and localization service
  • Integrated i18n module for static translation files management
  • Developed multi-language push notification system with notification groups in single API service
  • Managed match requests using Redis queue list type for efficient processing
  • Stored top matches in Redis using sorted lists and vector similarity scores for fast retrieval
  • Built chat media optimization service for image/video processing and delivery
  • Developed vector search API for semantic profile matching and similarity queries
  • Building Admin panel APIs for user, content, and report management

Achievements

  • Implemented single socket connection per user management preventing duplicate connections
  • Deployed Node.js cluster mode for horizontal scaling across multiple CPU cores
  • Configured Socket.io Redis adapter for cluster mode synchronization
  • Built SSE (Server-Sent Events) for general notifications with custom heartbeat and connection pooling
  • Optimized MongoDB aggregation pipelines with proper indexing and subdocuments to minimize lookups
  • Implemented patch writes in MongoDB for efficient partial updates
  • Created custom JWT guard with Passport including token blacklisting functionality
  • Handled cron jobs on single process using Redis distributed locks to prevent duplicate execution
  • Implemented data virtualization for admin panel using aggregated queries and pagination
  • Integrated Hugging Face models for profile embeddings improving match quality by 40%
  • Built hybrid matching system combining traditional filtering with AI-powered semantic similarity
  • Achieved 65% improvement in match relevance using vector embeddings vs traditional methods

Contributions

  • Adhered to TypeScript best practices using interfaces and type aliases throughout codebase
  • Implemented custom Redis service extending Redis IO for enhanced functionality including vector operations
  • Designed event-driven architecture for better service decoupling and maintainability
  • Created efficient Redis transport functionality for real-time features
  • Built scalable matching system handling high concurrent users with AI-powered recommendations
  • Implemented comprehensive multi-language support at all application layers
  • Optimized database queries and data structures for performance at scale
  • Designed robust real-time communication infrastructure with fallback mechanisms
  • Architected vector search integration using Redis Stack for semantic user matching
  • Implemented user profile tokenization pipeline using Hugging Face sentence-transformers
  • Built vector embedding generation service processing user bios, interests, and preferences
  • Created hybrid matching algorithm combining weighted preferences, cosine similarity, and vector search
  • Designed efficient vector indexing schema in Redis for fast K-nearest neighbor queries
  • Built automated reindexing pipeline updating user embeddings on profile changes
  • Implemented semantic similarity scoring combining traditional metrics with embedding distances
  • Optimized embedding model selection balancing accuracy vs inference speed for production use