🧠 NovaMind

The Operating System for Circular Intelligence. A shared, high-performance AI inference engine and model registry for the circular economy.

NovaMind (formerly NovaSapien) is the Horizontal Enabler responsible for Artificial Intelligence and Machine Learning. In the NovaEco "System of Systems," sectors should focus on domain logic, not MLOps.

NovaMind acts as the "Cortex" of the ecosystem. It provides a unified gRPC/REST interface for Computer Vision, Natural Language Understanding, and Predictive Analytics. Whether DurasAGV needs to identify a weed, or NovaRecycle needs to sort plastic, they both query NovaMind.

---

🎯 Value Proposition

Fragmented AI development leads to wasted compute and data silos. NovaMind centralizes intelligence:

1. Shared Wisdom: A "Plastic Identification Model" trained on data from a recycling plant in Berlin (NovaRecycle) is instantly available to a river cleanup robot in Bangkok (NovaNature).
2. Resource Efficiency: Centralized GPU clusters (or optimized Edge models) reduce the energy cost of inference, aligning with Green AI principles.
3. Privacy-Preserving: Supports Federated Learning, allowing factories (NovaMake) to train models on proprietary data without that data ever leaving their premises.

---

🏗️ Architecture (The Inference Hub)

NovaMind is built on NVIDIA Triton Inference Server to serve models via high-speed gRPC.

graph TD
    User((Sector Service)) -->|gRPC Request| API[NovaMind Gateway]
    
    subgraph "The Visual Cortex (Vision)"
        API -->|Image Frame| ResNet["ResNet-50 (Waste/Weeds)"]
        API -->|Satellite Tile| SegFormer["SegFormer (Forest Cover)"]
    end

    subgraph "The Auditory Cortex (Sound)"
        API -->|Audio Clip| BirdNET["BirdNET (Species ID)"]
        API -->|Machine Hum| Anomaly["Autoencoder (Predictive Maint)"]
    end

    subgraph "The Frontal Cortex (Logic)"
        API -->|Text Prompt| LLM["Mistral-7B (Circular Advisor)"]
        API -->|Time Series| XGBoost["XGBoost (Energy/Price Forecast)"]
    end

    subgraph "Model Registry"
        ResNet -.->|Fetch Weights| S3[("Model Storage")]
        LLM -.->|Fetch Weights| S3
    end

Loading

Integrated Consumers

• NovaRecycle: Streams camera feeds from conveyor belts. NovaMind returns coordinates of "PET Bottles" to robotic arms.
• NovaNature: Sends audio spectrograms. NovaMind returns "Species: Hylsa cinerea (Green Tree Frog)."
• NovaLab: Sends user intent text ("I have coffee grounds"). NovaMind uses LLMs to map this to "Mushroom Farming" business models.
• NovaEnergy: Sends weather forecasts. NovaMind predicts Solar PV output for the next 24 hours.

---

✨ Key Features

1. The Computer Vision Hub

Specialized models for circular tasks.

• WasteNet: A proprietary model trained on millions of images of crushed, dirty, and overlapping waste items (spectral + RGB).
• CropVision: Hyperspectral analysis for NovaAgro to detect nitrogen deficiency before it is visible to the human eye.

2. The Geometry Engine

For NovaMake and NovaBuild.

• Mesh Analysis: Checks 3D models for manifold errors (printability).
• Scan-to-BOM: Takes a 3D scan of a building and identifies "Standard Steel Beam IPE-300" for the material passport.

3. Predictive Maintenance (PdM)

Listening to the heartbeat of the infrastructure.

• NovaInfra sends vibration and audio data from pumps and turbines.
• NovaMind detects the specific frequency shift that precedes a bearing failure (Anomaly Detection), triggering a NovaSkills repair ticket.

4. Edge Sync

Not everything runs in the cloud.

• NovaMind manages the Model Distillation pipeline.
• It shrinks a 500MB cloud model into a 5MB quantized .tflite model and pushes it OTA (Over-the-Air) to DurasAGV robots for offline inference.

---

🚀 Getting Started

We use DevContainers with GPU support (NVIDIA Container Toolkit recommended).

Prerequisites

• Docker Desktop (Linux/WSL2 recommended for GPU passthrough)
• Python 3.11+
• 16GB+ RAM (for running quantized LLMs locally)

Installation

1. Clone the repo:

   git clone https://github.com/novaeco-tech/novamind.git
   cd novamind

2. Open in VS Code:
   • Run code .
   • Click "Reopen in Container" when prompted.
3. Start the Inference Server:

   make dev

   • Triton Metrics: http://localhost:8002/metrics
   • API Docs: http://localhost:8000/docs (FastAPI Wrapper)

Configuration (.env)

# Model Registry
MODEL_STORE_URI=s3://novaeco-models/production
DEFAULT_VISION_MODEL=wastenet-v4

# Hardware
ACCELERATOR=CUDA # or CPU, METAL (Mac)
BATCH_SIZE=32

---

📂 Repository Structure

novamind/
├── api/                # Python/FastAPI (The Gateway)
│   ├── src/
│   │   ├── routers/    # /vision, /audio, /text endpoints
│   │   └── pipeline/   # Pre/Post-processing logic (e.g., resizing images)
├── models/             # Model Configuration (config.pbtxt)
│   ├── wastenet/       # Waste classification configs
│   ├── birdnet/        # Bioacoustics configs
│   └── mistral/        # LLM parameter settings
├── training/           # Jupyter Notebooks & PyTorch Lightning scripts
│   └── experiments/    # Research sandbox
├── website/            # Documentation (Docusaurus)
└── tests/              # Model verification tests

---

🧪 Testing

We use Regression Testing for models to ensure "Smart" doesn't become "Stupid" after an update.

• Inference Test: make test-inference
  • Sends a reference image (sample_bottle.jpg) and asserts that the model returns "Class: Plastic" with >90% confidence.
• Latency Test: make test-perf
  • Measures the round-trip time for 1,000 requests. Ensures p99 latency is under 50ms for sorting applications.

---

🤝 Contributing

We need contributors with backgrounds in MLOps, Computer Vision, and Data Engineering. See CONTRIBUTING.md for details.

Maintainers: @novaeco-tech/maintainers-enabler-novamind