data/projects.json

[ {
    "id": "laptop-intel-engine",
    "title": "Cross-Marketplace Laptop & Review Intelligence Engine",
    "type": "AI/ML",
    "description": "Developed an advanced insights engine to analyse and compare laptops by integrating two distinct data sources: static technical specifications from PSREF PDFs and dynamic marketplace data (price, availability, reviews). Architected a multi-stage hybrid RAG system combining a FAISS vector index for semantic search and a SQLite database for structured SQL queries. Implemented query classification with guardrails to block or route non-RAG queries, reducing unnecessary compute and improving latency. Added contextual retrieval to enrich chunk representations, reducing top-20 retrieval failures by ~35%. Introduced a deep retrieval pipeline merging FAISS and BM25 with cross-encoder re-ranking for high-precision context selection. Added a query rewriter for improved retrieval accuracy and a query caching layer to minimize repeated embedding and LLM calls, lowering cost and enhancing response time.",
    "technologies": ["Python","FastAPI","Streamlit","Google Gemini API","FAISS","BM25","Sentence Transformers","Cross Encoders","SQLite","Pandas","Pydantic","Guardrails","Query Caching"
    ],
    "github_url": "https://github.com/tharu280",
    "demo_url": null,
    "image_path": "images/project_laptop.png"
  },
  {
    "id": "rag-book-recommender",
    "title": "RAG-based Advanced Book Recommendation System",
    "type": "AI/ML",
    "description": "Built a personalized book recommender using LangChain, HuggingFace Sentence Transformers, and ChromaDB, with zero-shot classification to auto-fill missing categories and enrich recommendations. Deployed the system using FastAPI, Docker, and Nginx on AWS Elastic Beanstalk for scalable, low-latency access.",
    "technologies": ["Python", "Numpy", "Pandas", "HuggingFace Transformers", "LangChain", "ChromaDB", "FastAPI", "Docker", "Nginx", "AWS Elastic Beanstalk"],
    "github_url": "https://github.com/tharu280",
    "demo_url": null,
    "image_path": "images/project_book.png"
  },
  {
    "id": "eyecon",
    "title": "EyeCon – Real-Time Blink Communication System",
    "type": "AI/ML",
    "description": "Developed the first system enabling fully paralysed users to communicate without any wearable devices, using only eye blinks detected via webcam. Implemented Eye Aspect Ratio (EAR)–based blink detection to differentiate short (next candidate) and long blinks (select candidate), achieving accurate, real-time interpretation of user intent. Integrated a Gemini LLM for context-aware candidate word suggestions, dynamically updating options as the sentence forms. Enabled automatic summarisation of composed sentences, providing users with coherent text output and facilitating streamlined communication.",
    "technologies": ["Python", "OpenCV", "MediaPipe", "Kivy", "Custom Morse Decoder", "Gemini LLM", "Contextual Candidate Selection", "FastAPI"],
    "github_url": "https://github.com/tharu280",
    "demo_url": null,
    "image_path": "images/project_eyecon.png"
  },
  {
    "id": "youtube-analytics",
    "title": "End-to-End YouTube Comment Analytics Pipeline",
    "type": "Data Engineering",
    "description": "Built an AI-powered pipeline to extract, classify, and visualize YouTube comments. Achieved 95.46% accuracy by fine-tuning BERT; also developed a BiLSTM classifier (90.75%); hosted the best model on HuggingFace Hub. Integrated Azure services (SQL Serverless, Data Factory, Power BI) for automated data flow, storage, and interactive visualisations.",
    "technologies": ["Python", "TensorFlow", "Keras", "TF-IDF", "BiLSTM", "BERT", "HuggingFace", "FastAPI", "Docker", "YouTube API", "Azure SQL Serverless", "Azure Data Factory", "Power BI", "Azure Data Studio"],
    "github_url": "https://github.com/tharu280",
    "demo_url": null,
    "image_path": "images/project_youtube.png"
  },
  {
    "id": "mental-health-chatbot",
    "title": "Comic Character-based RAG Mental Healthcare Assistant Chatbot",
    "type": "AI/ML",
    "description": "Built a RAG-based chatbot modelled after “Uncle Iroh” to provide mental health support using multi-source wisdom and personalized tone. Implemented 3 FAISS stores, with RunnableParallel & Sequential chains -cutting latency from 3.9s to 1.5s accurate and combined context retrieval. Optimised with Redis memory, LangSmith monitoring, and FastAPI; voice cloning via ElevenLabs in progress.",
    "technologies": ["Python", "LangChain", "HuggingFace Transformers", "Gemini 1.5 Flash", "FAISS", "Redis", "LangSmith", "ElevenLabs", "FastAPI", "Pydantic"],
    "github_url": "https://github.com/tharu280",
    "demo_url": null,
    "image_path": "images/project_chatbot.png"
  },
  {
    "id": "cafe-chatbot",
    "title": "AI-Powered RAG Chatbot for Cafe Business",
    "type": "Cloud AI",
    "description": "Built a Retrieval-Augmented Generation (RAG) chatbot using Amazon Bedrock, OpenSearch Serverless, and S3 to answer customer queries based on internal café documents. Preprocessed and embedded data for semantic search; integrated foundation models with prompt engineering for contextual responses. Designed a scalable, secure deployment with proper IAM configuration and AWS-native services to ensure real-world readiness.",
    "technologies": ["AWS Bedrock", "AWS OpenSearch Serverless", "AWS S3", "Llama 3 70B", "Titan Text G1 – Lite"],
    "github_url": "https://github.com/tharu280",
    "demo_url": null,
    "image_path": "images/project_cafe.png"
  },
  {
    "id": "ai-coding-agent",
    "title": "AI Coding Agent – V (Open Source Project)",
    "type": "Agentic AI",
    "description": "Designed and developed an AI coding agent demonstrating Agentic AI design patterns including Tool Use and Reflection Loop. Integrated Google Gemini’s function calling to perform grounded code operations such as reading, writing, and executing Python files. Implemented a 20-iteration reflection loop where the agent plans, acts, critiques, and improves results until completion.",
    "technologies": ["Python", "Google Gemini API", "Gemini SDK", "Pydantic", "Agentic design patterns"],
    "github_url": "https://github.com/tharu280",
    "demo_url": null,
    "image_path": "images/project_agent.png"
  },
  
  {
    "id": "food-ordering-backend",
    "title": "Event-Driven Backend for Food Ordering with Real-Time Fraud Detection",
    "type": "Backend/ML",
    "description": "Built a scalable event-driven backend using Python, Kafka, and Docker for a food ordering app with real-time processing. Modelled core events like order_placed, order_confirmed, and fraud_alert as Kafka topics across decoupled microservices. Integrated a fraud detection model into the transaction service to detect fraud in real-time and publish alerts via Kafka.",
    "technologies": ["Python", "Apache Kafka", "kafka-python", "scikit-learn", "Docker", "Pydantic", "MLflow", "FastAPI"],
    "github_url": "https://github.com/tharu280",
    "demo_url": null,
    "image_path": "images/project_food.png"
  },
  {
    "id": "tour-planner-agent",
    "title": "Tour Planner AI Agent",
    "type": "Agentic AI",
    "description": "Designed a hybrid AI system using LangGraph to orchestrate a multi-step agentic workflow for dynamic itinerary generation. Engineered a rule-based data ingestion pipeline to chain disparate APIs for gathering geocoding, route, and a broad set of location data. Implemented an “LLM as a Judge\" pattern that semantically analyses the user's qualitative query to rank and filter the raw data, delivering context-aware, personalized recommendations. Solved the 'Teleportartion bias' by a dynamic API process, Reducing the API requests.",
    "technologies": ["Python", "LangGraph", "FastAPI", "Google Gemini API", "Pydantic", "Geoapify", "OpenRouteService", "Nominatim"],
    "github_url": "https://github.com/tharu280",
    "demo_url": null,
    "image_path": "images/project_tour.png"
  }
]