Projects
Showcasing my work and contributions
MyMemory: AI-Powered Personal Memory Repository
Developed MyMemory, a production-grade personal memory repository, to enable users to easily store, search, and recall information using AI. Achieved 95%+ retrieval accuracy with a RAG pipeline employing Jina-AI for dense embeddings stored in Pinecone with metadata enrichment. Engineered real-time, WebSocket-powered chat with sub-2s responses using Groq-LLM. Deployed a scalable fullstack solution with frontend on Vercel and backend on Koyeb. Implemented modular code design with separate embedding, retrieval, and generation pipelines, and adopted scalable vector management using metadata filters and namespace isolation. Followed version-controlled development with environment-based secrets for secure deployments.
StreetSpecter: Real-Time Pothole Detection System
Designed and developed StreetSpecter, an AI-driven system for real-time pothole detection, aiming to improve road safety and maintenance efficiency. Custom-trained a YOLOv8s model on 5,493 images over 25 epochs using a Colab T4 GPU, achieving 90% detection accuracy. Integrated OpenStreetMap API to map over 1,000 km of road data, enabling geolocation of detected potholes and contributing to a potential 30% reduction in maintenance response times. Adopted best practices like modular Flask blueprints, route-based caching, and async-compatible I/O for OpenStreetMap data. Ensured reproducibility using configuration-driven training and automatic data augmentation pipelines. Deployed using Docker for consistent environment control across systems.
