mplads

Prompt

🛡️ AI-Powered Anomaly & Fraud Detection System for MPLADS Leveraging Machine Learning, NLP, and Graph Analytics to detect fund misuse, ghost projects, and financial anomalies in India's Member of Parliament Local Area Development Scheme — in real time. 📋 Table of Contents Problem Statement Objective Proposed Solution & Approach System Architecture Tech Stack Key Features Impact Future Scope Team Details Installation & Setup Screenshots & Demo License 🔴 Problem Statement What is MPLADS? The Member of Parliament Local Area Development Scheme (MPLADS) is a Government of India programme that allocates ₹5 crore annually per MP (both Lok Sabha and Rajya Sabha members) for developmental works in their constituencies — covering infrastructure like roads, schools, hospitals, water supply, and drainage. Since its inception in 1993, MPLADS has disbursed thousands of crores of public funds. While intended to empower MPs to address constituency-level needs, the scheme has been repeatedly cited in CAG (Comptroller and Auditor General) audit reports for serious financial and operational irregularities. The Real-World Problem Despite its intent, MPLADS suffers from systemic issues that erode public trust and drain taxpayer money: Issue Description 💸 Fund Misallocation Funds sanctioned for non-eligible works or outside constituency boundaries 🏚️ Ghost Projects Works reported as completed with no physical evidence on the ground 📄 Inflated Billing Contractors submitting invoices 2–5× above market benchmark rates 🔄 Duplicate Sanctions Same project approved multiple times under different work IDs ⏳ Delayed / Incomplete Works Projects stalled indefinitely with funds already released 🕳️ Vendor Collusion Single vendors cornering disproportionate contract shares across constituencies 🔇 Lack of Transparency Utilization certificates filed without independent verification According to CAG reports, a significant portion of audited MPLADS works showed procedural violations, cost overruns, or incomplete execution — yet payments were released without accountability. 🎯 Objective This system aims to automatically detect anomalies, fraud patterns, and operational inefficiencies within MPLADS data — shifting from reactive audits to proactive, AI-driven oversight. Who Benefits? Stakeholder Benefit 🏛️ Government Auditors (CAG/DPA) Automated flagging reduces manual audit workload by 70%+ 🏢 Ministry of Statistics & Programme Implementation Real-time compliance monitoring across all constituencies 🗺️ District Authorities Actionable alerts for delayed or suspicious works 📰 Journalists & RTI Activists Transparent, data-backed evidence of irregularities 👥 Citizens Public-facing accountability dashboard 🧠 Proposed Solution & Approach Data Sources MPLADS Portal Data — Work sanctions, fund releases, completion status, district-wise utilization reports Fund Utilization Reports — Payment installment records, UC (Utilization Certificate) filings Project Completion Records — Expected vs. actual completion timelines, geo-coordinates Geotagged Photographs — Site images tagged to GPS coordinates for physical verification Vendor Invoice Data — Contractor billing records, GST filings, registered vendor details Benchmark Rate Data — State PWD schedule of rates for cost comparison CAG Audit Reports — Historical flagged cases used as training signal AI / ML Techniques ┌─────────────────────────────────────────────────────────────────────────┐ │ AI/ML DETECTION LAYER │ ├───────────────────┬─────────────────────┬───────────────────────────────┤ │ Anomaly Detection│ NLP & Document AI │ Computer Vision │ │ ─────────────── │ ───────────────── │ ──────────────────────────── │ │ • Isolation │ • Invoice OCR & │ • Geotagged image analysis │ │ Forest │ entity extraction│ • Construction progress │ │ • DBSCAN / │ • Duplicate work │ estimation via CNNs │ │ LOF clustering │ detection (BERT │ • Ghost project detection │ │ • Autoencoder │ semantic sim.) │ (pre/post image diff) │ │ time-series │ • Inflated billing │ • GPS coordinate spoofing │ │ • Statistical │ keyword flagging │ detection │ │ Z-score / │ • Document │ │ │ Benford Law │ consistency NLP │ │ └───────────────────┴─────────────────────┴───────────────────────────────┘ ┌─────────────────────────────────────────────────────────────────────────┐ │ GRAPH ANALYTICS LAYER │ │ • Vendor–fund network graphs to detect concentration & shell entities │ │ • MP → Vendor → Contractor bipartite graph analysis │ │ • Community detection (Louvain) for collusion ring discovery │ │ • PageRank-based vendor risk scoring │ └─────────────────────────────────────────────────────────────────────────┘ Rule-Based + ML Hybrid Detection Logic The system combines hard rules (regulatory thresholds) with soft ML signals (statistical outliers) to produce a composite Risk Score (0.0–1.0): Detection Category Rule-Based Signal ML Signal Cost Overrun Cost > 120% of sanctioned amount Z-score vs. district/category peer group Timeline Violation Days elapsed > 2× expected duration Autoencoder reconstruction error on timeline Duplicate Work Same GPS ± 500m & same category BERT semantic similarity > 0.92 on title Vendor Concentration Single vendor > 40% of MP total works Graph centrality outlier detection Ghost Project No geotagged image within 6 months CV model: site activity score < threshold Billing Fraud Invoice > 2× PWD schedule rate Isolation Forest on unit cost distributions Unspent Funds >30% unspent after expected completion Time-series anomaly (LSTM forecasting) Final Risk Score = weighted combination of rule hits + ML anomaly scores, calibrated via historical CAG audit labels. 🏗️ System Architecture ┌──────────────────────────────────┐ │ DATA INGESTION │ │ MPLADS Portal · Invoices · │ │ Geotagged Images · PWD Rates │ └──────────────┬───────────────────┘ │ ▼ ┌──────────────────────────────────┐ │ PREPROCESSING & ETL │ │ Data cleaning · Normalisation · │ │ OCR pipeline · Image tagging │ └──────────────┬───────────────────┘ │ ┌──────────────▼───────────────────┐ │ ANOMALY DETECTION ENGINE │ │ ┌──────────┐ ┌──────────────┐ │ │ │ Rule │ │ ML Models │ │ │ │ Engine │ │ (ISO Forest,│ │ │ │ (Thres- │ │ LSTM, BERT, │ │ │ │ holds) │ │ Graph AI) │ │ │ └────┬─────┘ └──────┬───────┘ │ │ └────────┬───────┘ │ │ Risk Scorer │ └──────────────┬───────────────────┘ │ ┌──────────────▼───────────────────┐ │ REST API LAYER │ │ Flask · SQLite · JSON │ └──────────────┬───────────────────┘ │ ┌──────────────▼───────────────────┐ │ DASHBOARD & ALERT SYSTEM │ │ React · Interactive Maps · │ │ Risk Queues · Explainability │ └──────────────────────────────────┘ 💻 Tech Stack Backend Category Technology Language Python 3.11 API Framework Flask 3.x Database SQLite (production migration: PostgreSQL) ML Libraries scikit-learn, PyTorch, Transformers (HuggingFace), NetworkX NLP BERT, sentence-transformers, spaCy Computer Vision OpenCV, torchvision Analytics pandas, NumPy, SciPy Graph Analytics NetworkX, python-louvain Frontend Category Technology Framework React 18 + TypeScript Build Tool Vite Styling Vanilla CSS (custom design system) Charts Recharts Maps Leaflet.js / react-leaflet State Management React Hooks + Context Infrastructure Category Technology Dev Server Vite Dev + Flask (proxied) Report Generation Jinja2 templating + HTML export Version Control Git ✨ Key Features 🔴 Real-Time Fraud Scoring — Every MPLADS work item is assigned a composite risk score (0–100%) updated as new payment or completion data arrives, with color-coded severity badges (Critical / High / Medium / Low). 📊 MP-wise & District-wise Dashboards — Drill-down analytics across constituencies, districts, and states. Compare MPs against each other on compliance, spending velocity, and anomaly frequency. 🗺️ Anomaly Heatmaps — Geographic visualization of flagged works, overlaid on district maps to surface spatial clustering of fraud. ⚡ Priority Risk Queue — Auto-sorted queue of the highest-risk works requiring immediate audit attention, eliminating the need to manually sift through thousands of records. 🏢 Vendor Risk Profiling — Per-vendor risk scores, contract concentration metrics, and network graph of vendor–MP relationships to surface collusion patterns. 💡 Explainable AI Alerts — Every flagged case includes a human-readable explanation of why it was flagged (e.g., "Cost is 3.2× peer average in Rajasthan road works"), along with a recommended audit action. 📈 Benchmark Explorer — Compare any work's cost, timeline, or payment pattern against statistical benchmarks for similar works in the same terrain/category. 📋 Automated Report Generation — One-click PDF/HTML audit reports covering risk distributions, top anomalies, and compliance scores — ready for CAG or Ministry submission. 🔔 Alert & Notification Engine — Configurable threshold-based alerts for district collectors, state coordinators, and ministry officials. 🌟 Impact Impact Area Outcome 💰 Reduced Fund Leakage Early detection of billing fraud and ghost projects prevents crores in misappropriation per audit cycle ⏱️ Faster Audits Reduces manual audit effort from weeks to hours by surfacing only high-risk cases 🔍 Increased Transparency Public-facing dashboards enable citizen oversight and RTI-aligned reporting 🤝 Public Trust Demonstrates accountable, data-driven governance to citizens and media 📉 Deterrence Effect Visible AI monitoring deters misreporting at source 🏛️ Institutional Strengthening Provides CAG, DPA, and MOSPI with a scalable tool for concurrent audits 🚀 Future Scope 🔗 Blockchain-Based Fund Tracking — Immutable ledger for every MPLADS payment release, enabling tamper-proof audit trails 📡 Integration with RTI & CAG Systems — Direct API linkage for real-time data sharing with Right to Information portals and Comptroller General systems 📱 Mobile App for Ground-Level Verification — Field officers can submit geotagged photos and completion reports directly into the system 🌐 Scaling to Other Government Schemes — Plug-and-play architecture adaptable to PMGSY, MNREGS, Smart Cities Mission, AMRUT, and other central/state schemes 🤖 LLM-Powered Audit Chatbot — Natural language interface for auditors to query flagged cases 🛰️ Satellite Image Analysis — Automated construction-progress analysis using open satellite imagery (Sentinel-2) to verify physical project status 📊 Predictive Risk Modelling — Flag high-risk works before payment is released, using historical fraud signals and MP/vendor profiles 🏦 Cross-Scheme Deduplication — Detect the same physical location receiving funding from multiple schemes simultaneously 👥 Team Details Field Details Team Name (Your Team Name) Hackathon Smart India Hackathon (SIH) 2024 Problem Statement ID (PS ID) # Member Name Role 1 (Name) Team Lead / Full-Stack Developer 2 (Name) ML Engineer — Anomaly Detection 3 (Name) Data Engineer — ETL & Database 4 (Name) NLP / Computer Vision Engineer 5 (Name) Frontend Developer / UI Design 6 (Name) Backend Developer / API & Infra 🛠️ Installation & Setup Prerequisites Python 3.10+ with pip Node.js 18+ with npm Git 1. Clone the Repository git clone https://github.com/your-org/mplads-fraud-detection.git cd mplads-fraud-detection 2. Set Up the Backend cd backend # Create and activate virtual environment python -m venv venv # Windows venv\Scripts\activate # macOS / Linux source venv/bin/activate # Install dependencies pip install -r requirements.txt # Start the Flask API server python dashboard.py The backend API will be available at: http://localhost:5000 Health check: http://localhost:5000/api/summary 3. Set Up the Frontend Open a new terminal in the project root: cd frontend # Install Node dependencies npm install # Start the Vite development server npm run dev The frontend will be available at: http://localhost:5173 All /api/* requests are automatically proxied to the backend via vite.config.ts. 4. (Optional) Run the Analytics Engine To regenerate anomaly scores and risk flags from scratch: cd backend python analytics_engine.py python analytics_engine_phase3.py 5. (Optional) Generate Static HTML Report cd backend python generate_report.py # Output: mplads_report.html Environment Variables Create a .env file in backend/ if connecting to a live data source: DATABASE_URL=sqlite:///mplads.db FLASK_ENV=development FLASK_SECRET_KEY=your-secret-key-here 📸 Screenshots & Demo 🔗 Live Demo: (Add deployment URL here) 🎥 Demo Video: (Add YouTube / Google Drive link here) Screen Description Overview Dashboard MP-wise overview with risk distribution and priority queue Risk Cases View Flagged anomalies with explainability and risk scores Work Detail View Per-project deep dive with timeline, cost, and flag details Vendor Risk View Vendor concentration and network risk analysis Add screenshots to a screenshots/ folder in the project root and update the table above with image links. this is the mplads project we made for our sih hackathon. give me a list of questions that the judge could ask us based on this. questions and answers. ignore the computer vision part of the readme we didnt develop that