All Projects
Browse our complete collection of high-quality source codes.
SQL Assistant : Voice/Text to SQL Multilingual WebApp
INTRODUCTION - SQL Assistant is an AI-powered web application built with Streamlit that allows users to query any SQLite database using natural language — either by typing or speaking. The app uses Groq LLMs to convert plain English (or Hindi, Telugu, and 50+ other languages) into accurate SQL queries, executes them securely, and presents results with interactive charts, downloadable CSVs, and AI-generated summaries. Users can also bring their own database by uploading a .db file or CSV files directly from the sidebar. FEATURES - 1. Natural Language to SQL - Converts user questions into valid SQLite queries using Groq LLMs - Supports three models: Llama 3.3 70B Versatile, Llama 3.1 8B Instant, and Gemma 2 9B - Auto error correction with configurable reflection loop (0-10 retries) when the generated SQL is invalid 2. Voice Input with Multilingual Support - Built-in microphone recording using audio-recorder-streamlit - Speech-to-text powered by Groq Whisper (whisper-large-v3) - Automatically detects and transcribes 50+ languages including English, Hindi, and Telugu 3. Dynamic Database Management - Comes with a pre-built ecommerce SQLite database (customers, products, orders) - Upload any SQLite .db file from the sidebar to query it instantly - Upload one or more CSV files — each file automatically becomes a table in a new database - Schema is auto-detected from any database and passed dynamically to the LLM prompt - Sidebar schema viewer shows all tables, columns, types, and row counts 4. Interactive Plotly Visualizations - Four chart types: Bar Chart, Pie Chart, Line Chart, and Scatter Plot - Auto-selects appropriate columns (numeric vs categorical) for axes - Dark theme styled charts that match the app UI 5. SQL Security - SELECT-only enforcement — blocks INSERT, UPDATE, DELETE, DROP, ALTER, CREATE, TRUNCATE, and EXECUTE queries before they reach the database - Validation runs on every query before execution 6. CSV Export - One-click download button to export any query result as a .csv file 7. Performance Dashboard - Real-time tracking of total queries, successful queries, and failed queries - Success rate percentage with visual progress bar in the sidebar 8. AI-Powered Summarization - After every successful query, the LLM generates a natural language summary of the results - Users can add custom context (e.g., "write it in Hindi", "explain it simply") to control the summary style TECH STACK - - Frontend: Streamlit - Database: SQLite - LLM API: Groq (Llama 3.3 70B, Llama 3.1 8B, Gemma 2 9B) - Voice Transcription: Groq Whisper (whisper-large-v3) - Visualizations: Plotly Express - Audio Recording: audio-recorder-streamlit - SQL Formatting: sqlparse - Data Processing: pandas --> Response : 0.2 seconds --> Accuracy : 98.2%
Student Performance Prediction
This project implements a student performance prediction system using a modular machine learning pipeline built in Python. Nine regression models are trained and evaluated - Random Forest, XGBoost, CatBoost, Gradient Boosting, AdaBoost, Decision Tree, K-Nearest Neighbors, Linear Regression, and Extreme Learning Machine (ELM). ELM, introduced as the primary novel addition, is a Single Hidden Layer Feedforward Neural Network that analytically solves output weights via Moore-Penrose pseudoinverse, eliminating iterative backpropagation and achieving competitive accuracy at significantly faster training speeds (Deo et al., IEEE Access, 2020). All models are optimized using GridSearchCV and the best model by R² score is automatically saved for inference. A Flask web application provides a clean, responsive UI for real-time predictions with grade classification.
Intelligent File Organizer - ML
Upload any mix of files or folders and get back a neatly organized ZIP - automatically. FileGenie uses a RandomForest ML model to classify files by actual content (magic bytes, entropy, keywords), not just extension, so even mislabeled or extension-less files land in the right place. Built with FastAPI, React, and JWT authentication. Accuracy : 95% Tech Stack - Backend : Python 3.10+ · FastAPI · Uvicorn SQLAlchemy 2.0 ORM · SQLite (dev) / PostgreSQL (prod) PyJWT (HS256) scikit-learn RandomForestClassifier (300 trees, 16-feature vector) python-multipart for file uploads Frontend : React 18 · Vite 5 · Tailwind CSS 3 React Router v6 (multi-page SPA) JWT stored in localStorage · AuthContext for global auth state
Drowsiness Detection
This project detects drowsiness using classical computer vision techniques instead of deep learning. It first identifies the face using HOG + SVM (via dlib), then extracts 68 facial landmarks using an Ensemble of Regression Trees (ERT) model. From these landmarks, it calculates the Eye Aspect Ratio (EAR) to measure eye openness. If the EAR drops below a threshold (0.25) for a continuous number of frames, the system concludes that the eyes are closed for too long and triggers a drowsiness alert. App Modes - --> Live Webcam - Click Start to begin real-time detection - Click Stop to release the camera - Runs at ~30 FPS --> Upload Video - Supports `.mp4`, `.avi`, `.mov`, `.mkv` - Click **Run Detection** to process frame by frame - Progress bar shows completion --> Upload Image - Supports `.jpg`, `.jpeg`, `.png`, `.bmp`, `.webp` - Instant single-frame analysis - Shows annotated image with EAR value and status
Violence Classification System
Video Anomaly Detection (VAD) is an AI-powered violence classification system that analyzes video clips in real-time. Built on a MobileNetV2 + BiLSTM + Temporal Attention architecture, it detects and classifies four types of violent actions — Kick, Punch, Slap, and Group Violence — with up to 98% validation accuracy. The system extracts spatial features per frame using MobileNetV2, computes temporal difference features to capture motion dynamics, then passes the combined sequence through two Bidirectional LSTM layers and a Temporal Attention mechanism before classifying. Key highlights: Multi-clip inference (5 stride-based clips averaged) for robust prediction Focal loss training to prevent class collapse on minority classes Two-phase training strategy: feature learning then fine-tuning Streamlit web UI - upload any MP4/AVI/MOV file and get instant results
TRAVELYS – Smart Travel Risk & Cost Analytics System
TRAVELYS – Smart Travel Risk & Cost Analytics System TRAVELYS is a data-driven travel analytics and decision-support platform built with Flutter Web that helps users evaluate international travel destinations using multiple socioeconomic and environmental indicators. Instead of relying only on basic travel blogs or reviews, TRAVELYS aggregates global travel indexes, real-time weather data, currency exchange rates, and tourist attraction information to generate meaningful travel insights. The platform analyzes metrics such as safety, cost of living, quality of life, pollution, climate, and healthcare availability to help travelers make informed decisions before visiting a country. The application provides an interactive dashboard where users can explore country-level analytics, visualize travel-related indicators, and generate downloadable travel reports for deeper insights. Key Highlights: Data-Driven Travel Insights: Combines multiple global indicators including Quality of Life, Safety, Cost of Living, Healthcare, Climate, and Pollution to provide a comprehensive overview of travel conditions in different countries. Smart Travel Recommendation System: Generates a weighted travel recommendation score based on multiple indexes to help users identify destinations that best match their safety, affordability, and lifestyle preferences. Interactive Country Analytics Dashboard: Users can explore detailed country profiles displaying travel-related metrics, environmental conditions, and economic indicators in a clean and intuitive interface. Real-Time Travel Information: Integrates external APIs to fetch current weather conditions, currency exchange rates, and tourist attraction data for selected destinations. Tourist Attractions Explorer: Displays popular tourist destinations, landmarks, and points of interest using geographic data APIs to enhance travel planning. Downloadable Travel Reports: Users can generate downloadable reports summarizing travel insights, making it easier to save or share travel analysis. Share Travel Insights: Allows users to share country travel analytics through built-in sharing functionality. Secure User Authentication: User accounts are protected using Firebase Authentication with email verification to ensure secure access to the platform. Modern Responsive UI: Built using Flutter Web to provide a responsive, modern interface that works smoothly across desktops and browsers. Technology Stack Frontend: Flutter Web Dart Backend / Services: Firebase Authentication APIs Used: Geoapify API – Tourist attractions data Weather API – Real-time weather information Currency Exchange API – Global currency rates Data Sources: Global travel and lifestyle index datasets including: Quality of Life Index Purchasing Power Index Safety Index Health Care Index Cost of Living Index Traffic Commute Time Index Pollution Index Climate Index
Intrusion Detection
AI-Powered Network Intrusion Detection System An intelligent, real-time network intrusion detection system built with deep learning. It uses a CNN-LSTM hybrid neural network trained on the NSL-KDD benchmark dataset to automatically classify network traffic into five categories: Normal, DoS (Denial of Service), Probe, R2L (Remote-to-Local), and U2R (User-to-Root) attacks. The system features a sleek, modern web interface with a glassmorphic design, secure user authentication, and an interactive dashboard where users can input network traffic features and receive instant predictions with full confidence breakdowns across all threat classes. Key Highlights: - Deep Learning Powered: Multi-layer 1D Convolutional Neural Network combined with LSTM for temporal pattern recognition in network traffic, achieving 99% training accuracy. - Real-Time Threat Detection: Analyze network packets on the fly and receive predictions in under 100ms, enabling rapid threat identification and response. - 5-Class Attack Classification: Classifies traffic into DoS, Probe, R2L, U2R, or Normal — covering the most critical categories of network intrusions. - Secure Access: User registration and login with bcrypt-hashed passwords and session-based authentication to protect the dashboard and prediction endpoints. - Probability Breakdown: Every prediction comes with a full confidence distribution across all five threat classes, giving complete transparency into the model's decision. - Modern Web Dashboard: Beautiful, responsive UI with animated glassmorphism effects, interactive charts, and a clean user experience built with vanilla HTML, CSS, and JavaScript. - REST API: Programmatic access to the prediction engine via a JSON-based API endpoint for integration with external tools and automation pipelines. Technology Stack: - Backend: Python, Flask, TensorFlow/Keras - Frontend: HTML5, CSS3 (Glassmorphism), JavaScript - Machine Learning: CNN-LSTM Hybrid Model, Scikit-learn (preprocessing) - Database: SQLite3 (user authentication) - Dataset: NSL-KDD (Canadian Institute for Cybersecurity) - Security: bcrypt password hashing, session-based auth Built for cybersecurity researchers, network administrators, and anyone looking to leverage AI for proactive network defense.
Abusive Content Detection
Project Name: ToxicShield Description: A full-stack social media application containing a modern React frontend and a FastAPI backend with built-in AI toxicity moderation to ensure a safe, positive community environment. Key Features: - User Authentication: Secure sign-up and login interactions using JSON Web Tokens (JWT). - Social Feed: Users can create posts, upload image media, interact with the feed, and comment on others' posts. - Real-time AI Toxicity Moderation: All user-generated text is routed and analyzed in real-time by an advanced Deep Learning AI model (Detoxify/PyTorch) to detect toxic, obscene, threatening, or insulting language. - Automated Strike System: A zero-tolerance 3-strike moderation system automatically tracks and penalizes users who repeatedly violate community guidelines. - Admin Analytics Dashboard: View platform health, total warnings, and actively track the most toxic users. How the AI Moderation Works (Examples): 1. Innocent Content (Toxicity Score < 0.50) - Input Example: "I just had a wonderful day at the park!" - Result: The content is successfully published to the social feed. No strikes are added to the user's account. 2. Toxic Content Warning (Toxicity Score > 0.50) - Input Example: "Your opinion is ridiculous and you are dumb." - Result: The content is blocked from being published. The frontend displays a warning explicitly indicating a community guideline violation, and 1 strike point is added to their profile. 3. Severe Toxicity (Toxicity Score > 0.80) - Input Example: "You are an absolute piece of trash." - Result: The content is immediately blocked. The AI flags this as severe toxicity, and 3 strike points are instantly applied to the user's account. 4. Automated Account Suspension (Threshold >= 3 Strikes) - Condition: A user accumulates 3 or more strike points (either by 3 separate minor warnings or 1 severe toxicity violation). - Result: The user's account is automatically suspended for 7 days. They are forcefully logged out and blocked from logging back into the platform until the ban timeframe expires.
Ambulance Booking (Patient, Driver & Admin Dashboard)
A full-stack, real-time online ambulance booking and dispatch system. It connects patients with the nearest available ambulances within a 50km radius using GPS tracking. Technologies Used : Frontend : React Backend : Flask (Python) DB : SQLite3 Location Service : OpenStreetMap Features : Features include real-time location tracking on interactive maps, a 4-digit OTP verification system for secure patient pickups, driver dispatch notifications, and an automated fare/distance calculation upon trip completion. Important : The system supports three user roles: Patients (book rides), Drivers (accept/manage trips via a driver portal), and Hospital Admins (manage fleet and hospital data).
Marine Life & Waste Detection
An AI-powered underwater detection system for identifying marine species and underwater waste using computer vision and deep learning. Features : - Underwater Waste Detection - Detect 14+ types of underwater pollution (bottles, bags, nets, etc.) - Marine Species Detection - Identify sharks, turtles, rays, fish, and invertebrates - Video Detection - Process uploaded videos for detection - Live Analysis - Real-time detection from YouTube live streams and videos
Healthcare Patient Flow & Waiting Room Analytics
Healthcare Patient Flow & Waiting Room Analytics : PulseFlow Auditor is an advanced healthcare operational intelligence system that analyzes 171,064 (dataset cce licensed / kaggle) patient journeys to optimize hospital efficiency, predict capacity constraints, and quantify financial impact in Indian healthcare context. The system uses data-driven statistical analysis instead of arbitrary benchmarks to provide actionable insights for healthcare administrators. 🔧 Technical Stack : Backend: Python (Pandas, NumPy, SQLite3, PyWebView) Frontend: React 18 (Recharts, TailwindCSS, Vite) Database: SQLite with 171,064 patient journeys Architecture: Real-time Python-JavaScript bridge ⚡ Key Features : Advanced Analytics 9 Comprehensive KPIs: Operational, financial, and patient experience metrics Predictive Modeling: 24-hour volume forecasting with confidence intervals Real-Time Monitoring: Live alerts and capacity tracking Statistical Engine: Percentile-based thresholds (no arbitrary values)
Payment Fraud Detection (UPI & Cards)
Overview: UPI & Cards Payment Fraud Detection System This project is a standalone, AI-powered fraud detection system specifically designed for UPI (Unified Payments Interface) and card-based payment transactions. It leverages high-performance machine learning and a rule-based validation engine to identify potential fraudulent activity in real-time. Key Features : Real-Time Standalone Inference: Runs entirely within the browser using Pyodide (WebAssembly). This eliminates the need for a backend API for predictions, ensuring low latency and maximum privacy. Dual-Model Ensemble: Uses a weighted ensemble of XGBoost and LightGBM models, achieving a high degree of accuracy by combining the strengths of different gradient boosting architectures. UPI Validation Engine: Includes a specialized rule engine to catch structurally invalid Virtual Payment Addresses (VPAs) and Unique Transaction Reference (UTR) numbers specific to apps like GPay, PhonePe, and Paytm. Privacy-First History: Automatically saves transaction history to a local SQLite3 database (sql.js) stored within the browser's IndexedDB. No sensitive transaction data ever leaves your device. Advanced Feature Engineering: Automatically transforms raw transaction data into 447 engineered features, capturing complex temporal and statistical patterns found in real-world fraud. Comprehensive Metrics: Provides instant feedback with fraud probability scores (0–100%), risk levels (Low, Medium, High), and detailed performance insights. Technical Specifications: Model Performance: 98.45% Accuracy and 0.965 ROC-AUC, trained on over 590,000 real-world transactions (IEEE-CIS dataset). Frontend Stack: Built with React.js, Vite, and Tailwind CSS for a modern, responsive, and "Cyber-Fintech" user experience. Machine Learning: XGBoost (Binary Logistic) + LightGBM (GBDT) with weighted ensemble logic. Deployment: Fully client-side; can be served as a static site while maintaining full AI capabilities.
Parking Booking and Real Time Analysis
The frontend of the Smart Parking System is a responsive web application built with React and Vite. It provides an intuitive interface for users to interact with the parking predictor system. Key features include: - Real-Time Dashboard: Monitor live parking lot occupancy and slot statuses. - Interactive Map: Navigate to available parking slots using an integrated Leaflet map. - Slot Booking & Dynamic Pricing: Reserve parking slots with real-time price estimation based on current demand and time. - Occupancy Prediction Charts: View LSTM-based predictions for future parking occupancy trends using Chart.js.
PG & Mess Management - Client+Admin (MERN Stack)
PG Life is a comprehensive, MERN-stack based management system that streamlines Paying Guest House and Restaurant operations. It empowers administrators to efficiently manage rooms, mess plans, bookings, and support tickets, while providing a dedicated portal for students and users to request accommodations, report issues, and stay informed with announcements. Admin Features: - Room Management: Create, view, and delete room listings with details like size, price, bathroom status, and owner info. - Mess Management: Manage mess plans and daily menus. - Booking Control: Approve or cancel room and mess bookings. - Support Tickets: Respond to and close support tickets raised by residents. - Announcements: Post system-wide news and updates for all users. - Dashboard Overview: View key statistics like total users, estimated revenue, and open tickets. Client (Student/User) Features: - Room Browsing: Browse available rooms with filtering and search options. - Room Booking: Request room accommodations based on availability. - Mess Booking: Subscribe to mess plans for daily meals. - Support Tickets: Raise tickets for maintenance or any issues. - Announcements: Stay updated with the latest PG news and updates. - Reviews: Submit reviews and ratings for rooms and services.
Fake News Detection (Video, Audio, Text & Image)
FakeDetect is a cutting-edge, multimodal deepfake and fake news detection system designed for enterprise security and media verification. It uses state-of-the-art AI models to analyze Audio, Images, and Videos for signs of manipulation, synthesis, and deepfakes. Features: - Audio Detection: Analyzes wave patterns and voice anomalies to identify synthetic or cloned audio using a Wav2Vec2 sequence classifier. - Image Analysis: Detects AI-generated images, face swaps, and pixel-level manipulation using an EfficientNet V2 model. - Video Verification: Analyzes videos frame-by-frame and temporally using an Xception-based architecture. * Dynamic Heatmaps: Automatically generates temporal probability heatmaps showing the fake probability segment-by-segment. * Frame Extraction: Displays exactly which frames were analyzed by the engine for complete transparency. Tech Stack: - Backend: Python, FastAPI, TensorFlow/Keras (tf-keras), PyTorch, OpenCV, Librosa - Frontend: HTML5, Vanilla JS, CSS (Responsive, Modern UI) - Data Visualization: Matplotlib, Seaborn (for dynamic heatmaps) Overall Accuracy : 92%
Multi Model Cancer Detection
The Multimodal Cancer Diagnosis System is an advanced Neuro-Symbolic AI application designed to assist in the early detection and management of cancer by fusing data from three distinct sources: medical imaging (X-rays/CT scans), clinical notes (text), and patient vitals (structured data). Unlike traditional black-box AI models, this system combines deep learning (ResNet-18 for images, DistilBERT for text) with a deterministic rule-based engine, ensuring that critical medical rules (e.g., age risk factors, specific keywords) directly influence the final risk score for greater reliability and interpretability. The application features a modern, responsive interface with role-based access for both Patients and Doctors. Patients can upload reports for instant analysis, view a timeline of their medical history, receive personalized actionable recommendations (e.g., "Schedule Biopsy"), and interact with a Context-Aware Chatbot powered by Google Gemini that answers health queries using their specific medical records. Doctors have a dedicated dashboard to monitor patient risk scores and issue digital prescriptions, creating a comprehensive ecosystem for cancer care management. Accuracy : 90%+
Android Malware Detection
Android Malware Detection - A tool for quantitative risk analysis of Android applications based on machine learning techniques. Android Malware Detection is a tool for quantitative risk analysis of Android applications written in Java, which is used to check the permissions of the apps, and Python, which is used to compute a risk value based on apps' permissions. The tool uses classification techniques through scikit-learn, a machine learning library for Python, in order to generate a numeric risk value between 0 and 100 for a given app. In particular, the following classifiers of scikit-learn are used in Android Malware Detection (this list is chosen after extensive empirical assessments): * Support Vector Machines (SVM) * Multinomial Naive Bayes (MNB) * Gradient Boosting (GB) * Logistic Regression (LR) Unlike other tools, Android Malware Detection does not take into consideration only the permissions declared in the app manifest, but carries out reverse engineering on the apps to retrieve the bytecode and then infers through static analysis which permissions are actually used and which are not. In this way, it extracts four sets of permissions for every analyzed app: * Declared permissions: extracted from the app manifest * Exploited permissions: declared and actually used in the bytecode * Ghost permissions: not declared but with usages in the bytecode * Useless permissions: declared but never used in the bytecode
Disease Detection Using ML (Breast, Diabetes, Heart)
A machine learning-based application for detecting multiple diseases using clinical parameters. The system provides a unified interface to predict the likelihood of various health conditions. The app is working correctly with the three disease prediction models : ✅ Breast Cancer (92.98% accuracy) ✅ Diabetes (75.32% accuracy) ✅ Heart Disease (80.33% accuracy)
Use Gesture to Solve Maths
Hand Gesture Math Solver is an AI-powered computer vision project that allows users to solve mathematical expressions by drawing them in the air using hand gestures. The system captures real-time video from a webcam, tracks hand movements, and converts gestures into written mathematical expressions on a virtual canvas. Once the user submits the expression using a specific gesture, the drawing is sent to Google Gemini AI, which interprets and solves the math problem. The result is then displayed back in the application interface. This project demonstrates the powerful combination of Computer Vision, Gesture Recognition, and Generative AI, making it an innovative and interactive way to perform mathematical problem-solving.
Deepfake Detection (Only Video)
About the Project - The Deepfake Detection System is an AI-driven platform developed to identify manipulated digital media, focusing on videos. It integrates deep learning model to analyze visual artifacts and motion inconsistencies. The system is deployed through a Flask-based web application, providing users with authenticity scores, visual explanations, and secure access features. Key Points : - Detects fake videos and images using deep learning - Combines temporal and frame-level analysis for video deepfake detection - Uses ResNeXt50 + LSTM for motion inconsistency analysis in videos - Applies Xception network for frame-level artifact detection - Uses EfficientNet-B0 for image authenticity classification - Built with Flask backend and HTML/CSS/JavaScript frontend - Powered by PyTorch and TensorFlow/Keras frameworks - Provides fake probability scores and heatmaps for explainability - Includes a secure user authentication system Accuracy : ~96%
Deepfake Detection (Video, Image & Audio)
About This Project - The Deepfake Detection System is an AI-driven platform designed to identify manipulated digital media, including videos, images, and audio. It integrates multiple deep learning models to analyze visual artifacts, motion inconsistencies, and synthetic voice patterns. The system is deployed through a Flask-based web application that provides users with authenticity scores, visual explanations, and secure access features. Key Points : - Detects fake videos, images, and audio using advanced deep learning - Combines temporal and frame-level analysis for video deepfake detection - Uses ResNeXt50 + LSTM for motion inconsistency analysis in videos - Applies Xception Network for frame-level visual artifact detection - Uses EfficientNet-B0 for image authenticity classification - Integrates audio deepfake detection using spectrogram-based CNN models to analyze synthetic voice patterns - Built with a Flask backend and HTML/CSS/JavaScript frontend - Powered by PyTorch and TensorFlow/Keras frameworks - Provides fake probability scores, confidence graphs, and heatmaps for explainability - Includes a secure user authentication system Accuracy : ~96% overall detection accuracy across video, image, and audio deepfake datasets.
Skin Cancer Detection
Overview Skin Cancer Detection is an AI-powered medical imaging project designed to identify potential skin cancer from images of skin lesions. The system analyzes visual patterns in moles or abnormal skin growths and classifies them as benign (non-cancerous) or malignant (cancerous) using deep learning. This technology assists in early detection, which is critical because skin cancer especially melanoma can spread rapidly if not treated in time.
Deepfake Detection (Video & Image)
The Deepfake Detection System is an AI-driven platform developed to identify manipulated digital media, focusing on videos and images. It integrates multiple deep learning models to analyze visual artifacts and motion inconsistencies. The system is deployed through a Flask-based web application, providing users with authenticity scores, visual explanations, and secure access features. Key Points : - Detects fake videos and images using deep learning - Combines temporal and frame-level analysis for video deepfake detection - Uses ResNeXt50 + LSTM for motion inconsistency analysis in videos - Applies Xception network for frame-level artifact detection - Uses EfficientNet-B0 for image authenticity classification - Built with Flask backend and HTML/CSS/JavaScript frontend - Powered by PyTorch and TensorFlow/Keras frameworks - Provides fake probability scores and heatmaps for explainability - Includes a secure user authentication system Accuracy : ~96%