Real-time transformer fault monitoring powered by ensemble ML, quantum variational circuits, and web-grounded research. Built on NVIDIA infrastructure end-to-end.

┌─────────────────────────────────────────────────────────┐
│  Frontend (React + Vite)                                │
│  ├── WebSocket ← Real-time telemetry every 2s           │
│  ├── REST → /api/chat (Nemotron 4B via Ollama)          │
│  ├── REST → /api/search (Perplexity Sonar)              │
│  ├── REST → /api/predict (Quantum VQC)                  │
│  └── 3D CAD Visualizer (Three.js, 47 components)        │
└──────────────────────┬──────────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────────┐
│  FastAPI Backend (main.py :8000)                         │
│  ├── WebSocket /ws/telemetry (2s broadcast loop)         │
│  │                                                       │
│  ├── AI Pipeline                                         │
│  │   ├── ETT Risk Engine (XGBoost/LightGBM/CatBoost/RF) │
│  │   ├── DGA Fault Classifier (Rogers + Duval + VQC)     │
│  │   ├── Quantum VQC (6 qubits, 4 variational layers)   │
│  │   └── NemotronChat → Ollama :11434 (local GPU)        │
│  │                                                       │
│  ├── Integrations                                        │
│  │   └── PerplexityChat → Sonar API (web-grounded)       │
│  │                                                       │
│  └── 20 monitored transformers with full DGA profiles    │
└──────────────────────┬──────────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────────┐
│  Ollama (:11434)                                         │
│  └── nemotron-nano-4b-instruct (NVIDIA open model)       │
└─────────────────────────────────────────────────────────┘

# 1. Install dependencies and start the backend
cd backend && pip install -r requirements.txt
uvicorn main:app --host 0.0.0.0 --port 8000

# 2. Start the React frontend
cd ../frontend && npm install && npm run dev

# 3. (Optional) Enable local Nemotron chat
ollama pull nemotron-nano-4b-instruct && ollama serve

# 4. (Optional) Enable Perplexity web search
export PERPLEXITY_API_KEY=pplx-xxxx

All core monitoring features work without external API keys.

GridVeda uses a 4-stage pipeline that activates progressively based on risk severity.

ETT Risk Engine — Gradient-boosted ensemble (XGBoost, LightGBM, CatBoost, Random Forest) with 36 engineered features including thermal stress, Arrhenius aging factor, and joule heating proxy. Triggers DGA analysis when risk exceeds 50%.

DGA Fault Classifier — Consensus between Rogers Ratios (IEEE C57.104), Duval Triangle, and a trained ensemble. Classifies faults as Normal, Thermal, Discharge, or Arcing.

Quantum VQC — 6-qubit variational quantum circuit with 4 layers and 72 trainable parameters. Provides an independent validation signal through a fundamentally different computational paradigm. 3-way majority vote with Rogers and Duval methods.

Perplexity Sonar — Web-grounded research that finds real, documented transformer failures matching the detected fault pattern. Identifies the most likely failing physical component and highlights it on the 3D CAD model.

Full API documentation available at http://localhost:8000/docs.

• Transparency — All intermediate calculations (features, ratios, agreement scores, class probabilities) are exposed in every API response
• Consensus — No single model controls output; 3-way majority voting with automatic confidence penalties for disagreement
• Fairness — Physics-based deterministic scoring with identical thresholds across all transformers; no demographic data
• Safety — Tiered activation (ETT continuous, DGA at 50% risk, VQC alongside DGA); human-in-the-loop for all actions
• Privacy — All inference runs locally on NVIDIA hardware; sensor data never leaves the network