@afterlink/ai-assistant (Proprietary Neural Network Model)

A fully-custom, 100% offline Multi-Layer Perceptron (MLP) Neural Network built entirely from scratch in Node.js.

This package does not rely on Ollama, Llama, Gemini, OpenAI, or any other pre-trained third-party LLMs. It represents your own proprietary neural network trained directly on the AfterLink Communication Protocol source code, examples, and documentation, complete with high-fidelity CLI micro-animations.

Technical Architecture

• Input Layer: Vector size equal to vocabulary size (default: 800 most frequent tokens across the codebase). Evaluated using normalized bag-of-words.
• Hidden Layer: 64 neurons utilizing the ReLU (Rectified Linear Unit) activation function.
• Output Layer: Dynamic dimension matching the total number of knowledge categories and code blocks parsed (344 endpoints/specifications). Uses the Softmax activation function to output clear class probability distributions.
• Backpropagation Engine: Employs categorical cross-entropy loss with Stochastic Gradient Descent (SGD) weight updates.
• Inference Pipeline: Converts natural language queries into feature vectors using TF-IDF Inverse Document Frequency (IDF) scaling, runs the forward propagation pass, and retrieves the most statistically relevant codebase solutions.

Premium Console Micro-Animations

To provide a state-of-the-art interactive coding experience, the CLI terminal features custom physics-based TUI animations:

1. Sweep Text Shimmer: A dynamic, looping bright Cyan light sweep that runs across "🧠 Analyzing query against proprietary neural network..." during active neural reasoning, cleaning up seamlessly upon completion.
2. Line-by-Line Dot Matrix Scan: Phased console printing. Every output block—including code panels, boxes, headers, and resource lists—is first drawn as a sequence of high-contrast dot matrix loading placeholders (░░░░░░░) before scanning and revealing the actual formatted, colorized text.

Project Structure

packages/ai-assistant/
├── README.md               # User guide & documentation
├── package.json            # Scripts & project manifest
└── src/
    ├── afterlink-knowledge.json  # Ingested codebase database (generated)
    ├── vocabulary.json     # Trained vectorizer terms (generated)
    ├── model-weights.json  # Trained neural network weights (generated)
    ├── neural-net.js       # Forward-pass, Backprop & SGD Neural Network class
    ├── ingest.js           # Extracts text chunks from codebase
    ├── train.js            # Vocabulary compiler and Backprop training loop
    └── chat.js             # Interactive terminal UI executing local inferences and TUI animations

1. Prerequisites & Setup

Ensure you have Node.js 20.0.0 or higher installed. There are no other software dependencies (no Ollama, no API keys, no python binaries).

Install the package developer dependencies:

pnpm install

2. Training Your Custom Model

To index your codebase and train your proprietary neural network, execute the compilation pipeline:

# Step 1: Ingest documentation and source code
node packages/ai-assistant/src/ingest.js

# Step 2: Train the Multi-Layer Perceptron Neural Network
node packages/ai-assistant/src/train.js

During execution of train.js, the script will build the vectorizer vocabulary, calculate Inverse Document Frequencies (IDF), initialize the neural network matrices, perform gradient descent, and save the parameters.

3. Launching the Assistant

To query your trained model and seek assistance on building applications, troubleshooting bugs, or retrieving specifications with animated transitions:

node packages/ai-assistant/src/chat.js

Type a question such as:

• "How do I set up a secure AfterLink server with TLS?"
• "What is the structure of the 10-byte binary frame header?"
• "Explain the Zod schema validation rules."