🌌 Aetheris Neural Engine v2.1

A minimalist, high-performance Deep Learning inference node.

---

🛰️ Project Overview

Aetheris Neural Engine is a standalone desktop application designed for real-time handwritten digit recognition. Unlike standard "Hello World" AI projects, Aetheris implements a Deep Convolutional Neural Network (CNN) with a custom-built, cyber-industrial terminal interface.

The project bridges the gap between high-level cloud training (Google Colab) and local low-latency execution (Linux/Debian), providing a seamless "Draft-to-Deploy" workflow.

---

🛠️ Technical Architecture: The "Why"

1. The Brain: Convolutional Neural Network (CNN)

Instead of a simple Multi-Layer Perceptron (MLP), I chose a CNN architecture.

• Feature Extraction: By using Conv2d layers, the model learns spatial hierarchies of features (edges → curves → shapes) rather than treating pixels as independent variables.
• Batch Normalization: I implemented BatchNorm2d after convolutional layers to stabilize the learning process and significantly accelerate convergence (reaching ~99% accuracy in just 2 epochs).
• Dropout Regularization: To prevent overfitting, Dropout(0.5) is used in the fully connected layers, ensuring the engine generalizes well to "messy" human handwriting.

2. The Engine: PyTorch & AdamW

• PyTorch: Selected for its dynamic computation graph, essential for rapid prototyping and clean model definitions.
• AdamW Optimizer: I opted for AdamW (Adam with Weight Decay) instead of standard SGD. It provides better regularization and handles the training dynamics of deep networks more efficiently.

3. The Interface: PyWebView & Canvas API

• The Choice: Instead of outdated libraries like Tkinter, Aetheris uses PyWebView. This allows for a modern, hardware-accelerated UI using HTML5/CSS3.
• Zero-Latency Bridge: The communication between the JavaScript Canvas (where the user draws) and the Python Backend (where the model lives) is handled via a base64 encoded bi-directional bridge.

---

🧬 System Structure

The repository is structured for modularity and "Stateless" operation:

File	Purpose
src/engine.py	Contains the AetherisNet class. Keeping the architecture separate allows the model to be imported into other projects (e.g., NIB OS integration).
src/main.py	The "Control Room." Manages the window lifecycle, UI rendering, and the image-to-tensor preprocessing pipeline.
model/aetheris_model.pth	The serialized neural weights, optimized for CPU inference.

---

🚀 Installation & Deployment

1. Clone the Core

git clone https://github.com/NickIBrody/Aetheris-Neural-Engine cd Aetheris-Neural-Engine

2. Prepare the Environment (Linux/Debian)

Aetheris requires specific system-level engines to render its terminal UI

sudo apt update && sudo apt install -y
python3-gi python3-gi-cairo gir1.2-gtk-3.0 gir1.2-webkit2-4.1

3. Install Python Dependencies

pip install requirements.txt

4. Ignite the Engine

python3 src/main.py

🎮 MNIST Digit Classifier — Neural Core Interface

Hand-drawn digit recognition with a convolutional neural network.

🧠 Neural Core Init

Upon launch, the system verifies the model weights and initializes the graphical user interface (GUI).

✍️ Input Buffer

Use your mouse or touchpad to draw a digit (0–9) in the black central terminal zone.

🔍 Inference

Click the GUESS button. The image is:

• Captured from the canvas
• Resized to 28×28 pixels
• Normalized to MNIST standards
• Processed by the CNN

💬 Feedback

The AI responds in the chat interface with:

• Its predicted digit
• Confidence percentage (%)

---

📡 Future Roadmap

• Real-time Prediction
  Removing the "Guess" button for live, frame-by-frame inference.

• ONNX Export
  Converting the core to ONNX for an even lower memory footprint on edge devices.

• Dataset Expansion
  Fine-tuning the engine to recognize custom mathematical symbols.

---

---

---

Built with Python, PyTorch, and PyWebView.