🛡️ MPLADS AI Shield

AI-powered anomaly, compliance & risk intelligence for MPLADS project monitoring.

Turning fragmented records into explainable evidence, risk scores, and investigation priorities.

---

🚨 The Problem

MPLADS data is spread across recommendation, sanction, completion, payment and expenditure records, making manual monitoring difficult.

AI Shield helps identify:

• 🔎 Unusual project behaviour
• ⚠️ Compliance inconsistencies
• 💳 Unusual payment patterns
• 🔁 Similar or duplicate projects
• 📊 Projects requiring attention
• 🧠 Why a project was flagged

---

🧠 How It Works

Raw Data
    ↓
Preprocessing
    ↓
Canonical Projects
    ↓
Feature Engineering
    ↓
┌─────────────┬─────────────┬─────────────┐
│ Compliance  │  Anomalies  │  Duplicates │
└─────────────┴─────────────┴─────────────┘
        ↓
   Payment AI
        ↓
  Isolation Forest
        ↓
    Risk Fusion
        ↓
Risk Score + WHY Risky

---

🔬 AI / ML

• Compliance: Rule-based lifecycle & financial checks
• Anomaly Detection: Median/MAD, modified z-score & IQR
• Payment AI: Transaction and payment behaviour analysis
• Duplicate AI: Similarity and evidence-based matching
• Isolation Forest: Unsupervised anomaly detection
• Risk Fusion: Combines multiple evidence sources into a 0–100 score

Risk Weights

Evidence	Weight
Compliance	28
Financial Anomaly	20
Timeline Anomaly	12
Duplicate	12
Data Quality	8
Payment AI	10
Isolation Forest	10

---

Risk Fusion

Weighted evidence from: Compliance 28 | Financial 20 | Timeline 12 | Duplicate 12 | Data Quality 8 | Payment AI 10 | Isolation Forest 10

The current Phase 9 result is exposed by the authenticated GET /projects/{work_id}/risk endpoint in Phase 11. Legacy Phase 2 risk fields remain available on the existing project response for compatibility; they are not used as a fallback for the current Risk Fusion result.

📊 Current Results

43,863 projects analyzed

LOW        42,196
MEDIUM      1,662
HIGH            5
CRITICAL        0

---

🏗️ Tech Stack

Frontend: React, Vite, Tailwind CSS
Backend: FastAPI, Python, SQLAlchemy
Database: PostgreSQL
ML/Data: Pandas, NumPy, Scikit-learn

---

📁 Project Structure

MPLADS-AI-Shield/
├── backend/
│   ├── app/
│   ├── ml/
│   │   ├── preprocessing.py
│   │   ├── sources.py
│   │   ├── canonical.py
│   │   ├── features.py
│   │   ├── compliance/
│   │   ├── anomalies/
│   │   ├── duplicates/
│   │   ├── risk.py
│   │   └── risk_config.py
│   ├── data/
│   └── tests/
├── frontend/
│   └── src/
└── docs/

---

🚦 Project Status

✅ Phase 1 — Preprocessing
✅ Phase 2 — Canonicalization
✅ Phase 3 — Feature Engineering
✅ Phase 4 — Compliance
✅ Phase 5 — Anomaly Detection
✅ Phase 6 — Duplicate AI
✅ Phase 7 — Payment AI
✅ Phase 8 — Isolation Forest
✅ Phase 9 — Risk Fusion + WHY Risky

✅ Phase 10 — Evaluation + Synthetic Anomaly Tests
✅ Phase 11 — FastAPI ML Integration
🔜 Phase 12 — Upload & Analyze
🔜 Phase 13 — Role-Based Dashboard + Alerts
🔜 Phase 14 — SIH Polish & Demo

---

⚠️ Responsible AI

AI Shield identifies unusual patterns and projects requiring review.

It does not claim or prove fraud.

An anomaly is a signal for investigation, not proof of fraud.

---

🚀 Vision

Upload → Analyze → Detect → Explain → Prioritize → Investigate

From fragmented records to actionable intelligence.