@openfluke/welvet 0.1.2 → 0.1.4

package/README.md

@@ -1,731 +1,375 @@
 # @openfluke/welvet
 
-> TypeScript/JavaScript bindings for the LOOM neural network framework via WebAssembly
+Isomorphic TypeScript/JavaScript wrapper for the LOOM WebAssembly neural network framework.
 
-**Welvet** (Wrapper for Embedding Loom Via External Toolchain) provides a complete neural network API in the browser and Node.js/Bun environments. Built on WebAssembly, it delivers high-performance deep learning with zero dependencies.
+## Features
 
-[![npm version](https://img.shields.io/npm/v/@openfluke/welvet.svg)](https://www.npmjs.com/package/@openfluke/welvet)
-[![License](https://img.shields.io/badge/license-Apache%202.0-blue.svg)](LICENSE)
+- 🎉 **NEW: Simple API** - Streamlined functions with cross-platform consistency
+- 🚀 **Isomorphic WASM Wrapper** - Works in Node.js and browser with same API
+- 🔄 **Mirrors main.go** - Direct 1:1 mapping to WASM exports
+- 🎯 **Type-Safe** - Full TypeScript type definitions for all Network methods
+- 🤖 **Multi-Agent Networks** - Grid scatter architecture for heterogeneous agents
+- 📦 **JSON Configuration** - Build networks from simple JSON configs
+- ⚡ **Fast Training** - Optimized training with configurable parameters
+- 💾 **Model Persistence** - Save and load trained models as JSON
+- ✅ **Cross-Platform Consistency** - Same API as Python, C#, C, WASM
 
-## ✨ Features
-
-- 🚀 **5.4MB WASM Binary** - Complete neural network framework compiled to WebAssembly
-- 🧠 **All 5 Layer Types** - Dense, Conv2D, Multi-Head Attention, RNN, LSTM fully supported
-- 🎯 **Registry-based Initialization** - Dynamic layer creation via `CallLayerInit()` with zero manual exports
-- 🔍 **Runtime Introspection** - Discover methods, signatures, and parameters dynamically
-- 💾 **Model Serialization** - Save/load models as JSON (no filesystem required)
-- ⚡ **Full Training Support** - Train networks with `network.Train()` API and automatic gradients
-- 📘 **Full TypeScript Support** - Complete type definitions for IntelliSense
-- 🎯 **Zero Dependencies** - Pure WASM + Go runtime, no external libs
-- 🌐 **Isomorphic** - Works in browsers, Node.js, Bun, and Deno
-- 🎨 **Multiple Activation Functions** - ReLU, Sigmoid, Tanh, Softplus, LeakyReLU, Linear
-- ⚠️ **CPU-Only** (GPU support via WebGPU coming soon)
-
-## 📦 Installation
+## Installation
 
 ```bash
 npm install @openfluke/welvet
 ```
 
-Or with your preferred package manager:
-
-```bash
-# Yarn
-yarn add @openfluke/welvet
-
-# pnpm
-pnpm add @openfluke/welvet
-
-# Bun
-bun add @openfluke/welvet
-```
+## Quick Start
 
-## 🚀 Quick Start
+### 🎉 NEW: Simple API (Recommended)
 
-### The Easy Way: Load Complete Models
-
-Instead of manually configuring layers, **load a complete model with ONE line**:
+The simple API provides streamlined functions with consistent behavior across all platforms:
 
 ```typescript
-import { initLoom } from "@openfluke/welvet";
-
-const loom = await initLoom();
-
-// Load model from JSON (architecture + weights all at once!)
-const modelJSON = await fetch("test.json").then((r) => r.json());
-const network = loom.LoadModelFromString(
-  JSON.stringify(modelJSON),
-  "all_layers_test"
-);
+import {
+  init,
+  createNetworkFromJSON,
+  loadLoomNetwork,
+} from "@openfluke/welvet";
 
-// That's it! All 16 layers, weights, biases loaded automatically
-const input = new Array(10).fill(0).map(() => Math.random());
-const [output, duration] = JSON.parse(
-  network.ForwardCPU(JSON.stringify([input]))
-);
-console.log("Output:", output);
-```
+// Initialize LOOM WASM
+await init();
 
-**Live Demo:** See `wasm/all_layers_test.html` for a complete working example that loads a 26.4KB model with 16 layers (Dense, Conv2D, Attention, RNN, LSTM) and runs inference in the browser!
-
-### Manual Configuration (for building models from scratch)
-
-```typescript
-import { initLoom, ActivationType } from "@openfluke/welvet";
-
-// Initialize the WASM module
-const loom = await initLoom();
-
-// Create a neural network: 784 inputs → 392 hidden → 10 outputs
-const network = loom.NewNetwork(784, 1, 1, 2);
-
-// Configure layers using registry-based initialization
-const layer0 = loom.CallLayerInit(
-  "InitDenseLayer",
-  JSON.stringify([784, 392, ActivationType.ReLU])
-);
-const layer1 = loom.CallLayerInit(
-  "InitDenseLayer",
-  JSON.stringify([392, 10, ActivationType.Sigmoid])
-);
-
-network.SetLayer(JSON.stringify([0, 0, 0, JSON.parse(layer0)]));
-network.SetLayer(JSON.stringify([0, 0, 1, JSON.parse(layer1)]));
-
-// Forward pass
-const input = new Array(784).fill(0).map(() => Math.random());
-const resultJSON = network.ForwardCPU(JSON.stringify([input]));
-const [output, duration] = JSON.parse(resultJSON);
+// Create network from JSON config
+const config = {
+  batch_size: 1,
+  grid_rows: 1,
+  grid_cols: 3,
+  layers_per_cell: 1,
+  layers: [
+    { type: "dense", input_size: 8, output_size: 16, activation: "relu" },
+    {
+      type: "parallel",
+      combine_mode: "grid_scatter",
+      grid_output_rows: 3,
+      grid_output_cols: 1,
+      grid_output_layers: 1,
+      grid_positions: [
+        { branch_index: 0, target_row: 0, target_col: 0, target_layer: 0 },
+        { branch_index: 1, target_row: 1, target_col: 0, target_layer: 0 },
+        { branch_index: 2, target_row: 2, target_col: 0, target_layer: 0 },
+      ],
+      branches: [
+        {
+          type: "parallel",
+          combine_mode: "add",
+          branches: [
+            {
+              type: "dense",
+              input_size: 16,
+              output_size: 8,
+              activation: "relu",
+            },
+            {
+              type: "dense",
+              input_size: 16,
+              output_size: 8,
+              activation: "gelu",
+            },
+          ],
+        },
+        { type: "lstm", input_size: 16, hidden_size: 8, seq_length: 1 },
+        { type: "rnn", input_size: 16, hidden_size: 8, seq_length: 1 },
+      ],
+    },
+    { type: "dense", input_size: 24, output_size: 2, activation: "sigmoid" },
+  ],
+};
 
-console.log("Output:", output);
-console.log("Inference time:", duration / 1e6, "ms");
+const network = createNetworkFromJSON(JSON.stringify(config));
 
-// Training with the high-level API
+// Training
 const batches = [
-  { Input: input, Target: [0, 0, 0, 0, 0, 0, 0, 0, 0, 1] }, // Example: classify as "9"
+  { Input: [0.2, 0.2, 0.2, 0.2, 0.8, 0.8, 0.8, 0.8], Target: [1.0, 0.0] },
+  { Input: [0.9, 0.9, 0.9, 0.9, 0.1, 0.1, 0.1, 0.1], Target: [0.0, 1.0] },
+  { Input: [0.7, 0.7, 0.7, 0.7, 0.3, 0.3, 0.3, 0.3], Target: [0.0, 1.0] },
+  { Input: [0.3, 0.3, 0.3, 0.3, 0.7, 0.7, 0.7, 0.7], Target: [1.0, 0.0] },
 ];
-const config = {
-  Epochs: 10,
-  LearningRate: 0.01,
+
+const trainingConfig = {
+  Epochs: 800,
+  LearningRate: 0.15,
+  UseGPU: false,
+  PrintEveryBatch: 0,
   GradientClip: 1.0,
   LossType: "mse",
+  Verbose: false,
 };
 
-const trainingResult = network.Train(JSON.stringify([batches, config]));
-const result = JSON.parse(trainingResult);
-console.log("Final loss:", result.FinalLoss);
-```
-
-## 📚 API Reference
-
-### Initialization
-
-```typescript
-interface InitOptions {
-  wasmUrl?: string | URL;      // Custom WASM file location
-  injectGoRuntime?: boolean;   // Include Go runtime (default: true)
-}
-
-const loom = await initLoom(options?);
-```
-
-### Creating Networks
-
-```typescript
-const network = loom.NewNetwork(
-  inputSize: number,     // Input layer size
-  gridRows: number,      // Grid rows (use 1 for simple networks)
-  gridCols: number,      // Grid columns (use 1 for simple networks)
-  layersPerCell: number  // Number of layers
-);
-```
-
-### Layer Types
-
-All layer types are created via the registry system using `CallLayerInit()`:
-
-#### Dense (Fully-Connected) Layer
-
-```typescript
-const config = loom.CallLayerInit(
-  "InitDenseLayer",
-  JSON.stringify([
-    inputSize: number,
-    outputSize: number,
-    activation: ActivationType,
-  ])
-);
-```
-
-#### Conv2D Layer
+const [result] = network.Train(JSON.stringify([batches, trainingConfig]));
+console.log("Training complete!");
 
-```typescript
-const config = loom.CallLayerInit(
-  "InitConv2DLayer",
-  JSON.stringify([
-    height: number,        // Input height
-    width: number,         // Input width
-    channels: number,      // Input channels
-    filters: number,       // Number of output filters
-    kernelSize: number,    // Kernel size (e.g., 3 for 3x3)
-    stride: number,        // Stride (typically 1 or 2)
-    padding: number,       // Padding (typically 0 or 1)
-    activation: ActivationType,
-  ])
+// Forward pass
+const [output] = network.ForwardCPU(
+  JSON.stringify([[0.2, 0.2, 0.2, 0.2, 0.8, 0.8, 0.8, 0.8]])
 );
-```
-
-#### Multi-Head Attention Layer
-
-```typescript
-const config = loom.CallLayerInit(
-  "InitMultiHeadAttentionLayer",
-  JSON.stringify([
-    seqLength: number,     // Sequence length
-    dModel: number,        // Model dimension
-    numHeads: number,      // Number of attention heads
-    activation: ActivationType,
-  ])
+console.log("Output:", JSON.parse(output)); // [0.950, 0.050]
+
+// Evaluate network
+const inputs = batches.map((b) => b.Input);
+const expected = [0, 1, 1, 0];
+const [metrics] = network.EvaluateNetwork(JSON.stringify([inputs, expected]));
+const metricsData = JSON.parse(metrics);
+console.log(
+  `Quality: ${metricsData.score}/100, Deviation: ${metricsData.avg_deviation}%`
 );
-```
 
-#### RNN Layer
+// Save/Load
+const [modelJSON] = network.SaveModelToString(JSON.stringify(["my_model"]));
+console.log(`Model saved (${modelJSON.length} bytes)`);
 
-```typescript
-const config = loom.CallLayerInit(
-  "InitRNNLayer",
-  JSON.stringify([
-    inputSize: number,     // Input feature size
-    hiddenSize: number,    // Hidden state size
-    seqLength: number,     // Sequence length
-    outputSize: number,    // Output size (hiddenSize * seqLength)
-  ])
-);
-```
-
-#### LSTM Layer
-
-```typescript
-const config = loom.CallLayerInit(
-  "InitLSTMLayer",
-  JSON.stringify([
-    inputSize: number,     // Input feature size
-    hiddenSize: number,    // Hidden/cell state size
-    seqLength: number,     // Sequence length
-    outputSize: number,    // Output size (hiddenSize * seqLength)
-  ])
+// Load model
+const loadedNetwork = loadLoomNetwork(modelJSON, "my_model");
+const [output2] = loadedNetwork.ForwardCPU(
+  JSON.stringify([[0.2, 0.2, 0.2, 0.2, 0.8, 0.8, 0.8, 0.8]])
 );
-```
-
-**Activation Types:**
-
-```typescript
-enum ActivationType {
-  ReLU = 0, // Scaled ReLU (1.1x)
-  Sigmoid = 1, // Logistic sigmoid
-  Tanh = 2, // Hyperbolic tangent
-  Softplus = 3, // Smooth ReLU
-  LeakyReLU = 4, // ReLU with 0.1x negative slope
-  Linear = 5, // Identity (no activation)
-}
-```
-
-activation: ActivationType // ReLU, Sigmoid, Tanh, Linear
-);
-
-network.SetLayer(JSON.stringify([
-gridRow, // Grid row index
-gridCol, // Grid column index
-layerIndex, // Layer index within cell
-JSON.parse(config)
-]));
+// output2 === output (bit-for-bit identical!)
+```
+
+**Simple API Functions:**
+
+- `createNetworkFromJSON(jsonConfig)` - Create network from JSON
+- `loadLoomNetwork(jsonString, modelID)` - Load saved model
+- `network.ForwardCPU(inputJSON)` - Forward pass
+- `network.BackwardCPU(gradientsJSON)` - Backward pass
+- `network.Train(paramsJSON)` - Train network
+- `network.SaveModelToString(idJSON)` - Save to JSON string
+- `network.EvaluateNetwork(paramsJSON)` - Evaluate with metrics
+- `network.UpdateWeights(lrJSON)` - Update weights
+
+**Cross-Platform Results:**
+
+- ✅ Same training: 99.5% improvement, 100/100 quality score
+- ✅ Same save/load: 0.00 difference in predictions
+- ✅ Same evaluation: Identical deviation metrics
+- ✅ Same behavior as Python, C#, C, and WASM
+
+See `example/grid-scatter.ts` for a complete working example.
+{
+type: "parallel",
+combine_mode: "add",
+branches: [
+{
+type: "dense",
+input_size: 16,
+output_size: 8,
+activation: "relu",
+},
+{
+type: "dense",
+input_size: 16,
+output_size: 8,
+activation: "gelu",
+},
+],
+},
+{ type: "lstm", input_size: 16, hidden_size: 8, seq_length: 1 },
+{ type: "rnn", input_size: 16, hidden_size: 8, seq_length: 1 },
+],
+},
+{
+type: "dense",
+input_size: 24,
+output_size: 2,
+activation: "sigmoid",
+},
+],
+});
 
-````
+// Train multi-agent network
+const batches: TrainingBatch[] = [
+{ Input: [0.2, 0.2, 0.2, 0.2, 0.8, 0.8, 0.8, 0.8], Target: [1.0, 0.0] },
+{ Input: [0.9, 0.9, 0.9, 0.9, 0.1, 0.1, 0.1, 0.1], Target: [0.0, 1.0] },
+];
 
-#### Multi-Head Attention Layer
+const config: TrainingConfig = {
+Epochs: 800,
+LearningRate: 0.15,
+LossType: "mse",
+Verbose: false,
+};
 
-```typescript
-const config = loom.InitMultiHeadAttentionLayer(
-  dModel: number,       // Model dimension
-  numHeads: number,     // Number of attention heads
-  seqLength: number,    // Sequence length
-  activation: ActivationType
-);
+const result = agentNetwork.Train(JSON.stringify([batches, config]));
 
-network.SetLayer(JSON.stringify([0, 0, layerIndex, JSON.parse(config)]));
 ````
 
-### Training Operations
-
-#### Forward Pass
-
-```typescript
-const input = [0.1, 0.2, 0.3, 0.4];
-const resultJSON = network.ForwardCPU(JSON.stringify([input]));
-const [output, duration] = JSON.parse(resultJSON);
-```
-
-#### Backward Pass
-
-```typescript
-const gradOutput = new Array(outputSize).fill(0.01);
-const backwardJSON = network.BackwardCPU(JSON.stringify([gradOutput]));
-const [gradInput, duration] = JSON.parse(backwardJSON);
-```
-
-#### Update Weights
+## API Reference
 
-```typescript
-const learningRate = 0.01;
-network.UpdateWeights(JSON.stringify([learningRate]));
-```
-
-### Model Persistence
+### Functions
 
-#### Load Model (The Easy Way - ONE LINE!)
+#### `async init(): Promise<void>`
 
-```typescript
-// Fetch model from server
-const savedModel = await fetch("model.json").then((r) => r.json());
-
-// Load complete network with ONE function call!
-const network = loom.LoadModelFromString(
-  JSON.stringify(savedModel),
-  "model_name"
-);
+Initialize LOOM WASM module for Node.js environment.
 
-// Or from localStorage
-const savedModel = JSON.parse(localStorage.getItem("my_model")!);
-const network = loom.LoadModelFromString(
-  JSON.stringify(savedModel),
-  "model_name"
-);
-```
+#### `async initBrowser(): Promise<void>`
 
-**That's it!** All layers, weights, biases, and configurations are automatically restored. No manual layer setup needed!
+Initialize LOOM WASM module for browser environment.
 
-#### Save Model
+#### `createNetwork(config: object | string): Network`
 
-```typescript
-const modelJSON = network.SaveModelToString(JSON.stringify(["model_name"]));
-const model = JSON.parse(JSON.parse(modelJSON)[0]);
+Create a new neural network from JSON configuration object or string.
 
-// Store anywhere (localStorage, IndexedDB, etc.)
-localStorage.setItem("my_model", JSON.stringify(model));
-```
+**Note:** This is the only global function exposed by the WASM (mirrors `createLoomNetwork` from main.go). To load a saved model, just pass the saved JSON string to `createNetwork()`.
 
-#### Save Model
+### Network Interface
 
-```typescript
-const modelJSON = network.SaveModelToString(JSON.stringify(["model_name"]));
-const model = JSON.parse(JSON.parse(modelJSON)[0]);
+The `Network` object returned by `createNetwork()` has all methods from the Go `nn.Network` type automatically exposed via reflection.
 
-// Store anywhere (localStorage, IndexedDB, backend API, etc.)
-localStorage.setItem("my_model", JSON.stringify(model));
-```
+**Important:** All Network methods follow the WASM calling convention:
 
-#### Cross-Platform Model Loading
+- Take a single parameter: JSON string of an array of parameters
+- Return a JSON string of an array of results
 
-The same JSON model file works across **all three platforms**:
+Example:
 
 ```typescript
-// JavaScript/WASM
-const network = loom.LoadModelFromString(modelJSON, "model_id");
-```
-
-```python
-# Python
-network = welvet.load_model_from_string(model_json, "model_id")
-```
+// Method with no parameters
+const info = network.GetNetworkInfo(JSON.stringify([]));
+const parsed = JSON.parse(info)[0];
 
-```go
-// Go
-network, _ := nn.LoadModelFromString(modelJSON, "model_id")
-```
+// Method with parameters
+const result = network.Train(JSON.stringify([batches, config]));
+const data = JSON.parse(result)[0];
 
-See `examples/all_layers_validation.go` for a complete demo that generates test.json (26.4KB with 16 layers) and verifies all three platforms load it identically!
-
-#### Load Model (Legacy API)
-
-````
-
-### Runtime Introspection
-
-#### Get All Methods
-
-```typescript
-const methodsJSON = network.GetMethods();
-const methods = JSON.parse(methodsJSON);
-
-methods.forEach((method) => {
-  console.log(
-    `${method.method_name}(${method.parameters.map((p) => p.type).join(", ")})`
-  );
-});
+// Save model (requires modelID parameter)
+const saved = network.SaveModelToString(JSON.stringify(["my-model"]));
+const json = JSON.parse(saved)[0];
 ````
 
-#### Check Method Availability
+#### Available Network Methods
+
+- `ForwardCPU(paramsJSON)` - CPU forward pass: `[inputs]`
+- `ForwardGPU(paramsJSON)` - GPU forward pass: `[inputs]`
+- `BackwardCPU(paramsJSON)` - CPU backward pass: `[gradients]`
+- `BackwardGPU(paramsJSON)` - GPU backward pass: `[gradients]`
+- `UpdateWeights(paramsJSON)` - Update weights: `[learningRate]`
+- `Train(paramsJSON)` - Train network: `[batches, config]`
+- `SaveModelToString(paramsJSON)` - Save model: `["modelID"]`
+- `GetWeights(paramsJSON)` - Get layer weights: `[row, col, layer]`
+- `SetWeights(paramsJSON)` - Set layer weights: `[row, col, layer, weights]`
+- `GetBiases(paramsJSON)` - Get layer biases: `[row, col, layer]`
+- `SetBiases(paramsJSON)` - Set layer biases: `[row, col, layer, biases]`
+- `GetActivation(paramsJSON)` - Get activation: `[row, col, layer]`
+- `GetLayerType(paramsJSON)` - Get layer type: `[row, col, layer]`
+- `GetLayerSizes(paramsJSON)` - Get layer sizes: `[row, col, layer]`
+- `GetBatchSize(paramsJSON)` - Get batch size: `[]`
+- `GetGridDimensions(paramsJSON)` - Get grid dimensions: `[]`
+- `GetNetworkInfo(paramsJSON)` - Get network info: `[]`
+- `GetTotalParameters(paramsJSON)` - Get parameter count: `[]`
+- `InitializeWeights(paramsJSON)` - Initialize weights: `[]` or `[method]`
+- `Clone(paramsJSON)` - Clone network: `[]`
+- And 10+ more methods...
+- `GetLastOutput(): string` - Get last forward pass output
+
+### Types
+
+#### `NetworkConfig`
 
 ```typescript
-if (network.HasMethod("ForwardGPU")) {
-  const signature = network.GetMethodSignature(JSON.stringify(["ForwardGPU"]));
-  console.log(signature);
+interface NetworkConfig {
+  batch_size: number;
+  grid_rows?: number; // Required for grid networks (use 1 for sequential)
+  grid_cols?: number; // Required for grid networks (use 1 for sequential)
+  layers_per_cell?: number; // Required for grid networks
+  layers: LayerConfig[];
 }
 ```
 
-#### List Method Names
+#### `LayerConfig`
 
 ```typescript
-const names = JSON.parse(network.ListMethods());
-console.log("Available methods:", names);
-```
-
-## 🎨 Activation Functions
-
-```typescript
-enum ActivationType {
-  ReLU = 0, // Rectified Linear Unit
-  Sigmoid = 1, // Sigmoid (logistic)
-  Tanh = 2, // Hyperbolic tangent
-  Linear = 3, // No activation (identity)
+interface LayerConfig {
+  type: string;
+  input_size?: number;
+  output_size?: number;
+  hidden_size?: number;
+  seq_length?: number;
+  activation?: string;
+  combine_mode?: string;
+  grid_output_rows?: number;
+  grid_output_cols?: number;
+  grid_output_layers?: number;
+  grid_positions?: GridPosition[];
+  branches?: LayerConfig[];
 }
 ```
 
-## 💡 Complete Examples
-
-### MNIST-Style Classifier
+#### `TrainingBatch`
 
 ```typescript
-import { initLoom, ActivationType } from "@openfluke/welvet";
-
-async function trainMNIST() {
-  const loom = await initLoom();
-
-  // Network: 784 → 128 → 64 → 10
-  const network = loom.NewNetwork(784, 1, 1, 3);
-
-  const layer0 = loom.InitDenseLayer(784, 128, ActivationType.ReLU);
-  const layer1 = loom.InitDenseLayer(128, 64, ActivationType.ReLU);
-  const layer2 = loom.InitDenseLayer(64, 10, ActivationType.Sigmoid);
-
-  network.SetLayer(JSON.stringify([0, 0, 0, JSON.parse(layer0)]));
-  network.SetLayer(JSON.stringify([0, 0, 1, JSON.parse(layer1)]));
-  network.SetLayer(JSON.stringify([0, 0, 2, JSON.parse(layer2)]));
-
-  // Training loop
-  const epochs = 50;
-  const learningRate = 0.01;
-
-  for (let epoch = 0; epoch < epochs; epoch++) {
-    // Your training data here
-    const input = new Array(784).fill(0).map(() => Math.random());
-    const target = new Array(10).fill(0);
-    target[Math.floor(Math.random() * 10)] = 1;
-
-    // Forward
-    const [output] = JSON.parse(network.ForwardCPU(JSON.stringify([input])));
-
-    // Compute loss (MSE)
-    const loss =
-      output.reduce((sum, val, i) => sum + Math.pow(val - target[i], 2), 0) /
-      output.length;
-
-    // Backward
-    const gradOutput = output.map(
-      (val, i) => (2 * (val - target[i])) / output.length
-    );
-    network.BackwardCPU(JSON.stringify([gradOutput]));
-
-    // Update
-    network.UpdateWeights(JSON.stringify([learningRate]));
-
-    if (epoch % 10 === 0) {
-      console.log(`Epoch ${epoch}: Loss = ${loss.toFixed(6)}`);
-    }
-  }
-
-  // Save model
-  const modelJSON = network.SaveModelToString(JSON.stringify(["mnist"]));
-  localStorage.setItem("mnist_model", JSON.parse(modelJSON)[0]);
+interface TrainingBatch {
+  Input: number[];
+  Target: number[];
 }
 ```
 
-### XOR Problem
+#### `TrainingConfig`
 
 ```typescript
-import { initLoom, ActivationType } from "@openfluke/welvet";
-
-const loom = await initLoom();
-const network = loom.NewNetwork(2, 1, 1, 2);
-
-// 2 → 4 → 1 (XOR needs hidden layer)
-const layer0 = loom.InitDenseLayer(2, 4, ActivationType.ReLU);
-const layer1 = loom.InitDenseLayer(4, 1, ActivationType.Sigmoid);
-
-network.SetLayer(JSON.stringify([0, 0, 0, JSON.parse(layer0)]));
-network.SetLayer(JSON.stringify([0, 0, 1, JSON.parse(layer1)]));
-
-const trainingData = [
-  { input: [0, 0], target: [0] },
-  { input: [0, 1], target: [1] },
-  { input: [1, 0], target: [1] },
-  { input: [1, 1], target: [0] },
-];
-
-for (let epoch = 0; epoch < 1000; epoch++) {
-  let totalLoss = 0;
-
-  for (const sample of trainingData) {
-    const [output] = JSON.parse(
-      network.ForwardCPU(JSON.stringify([sample.input]))
-    );
-    const loss = Math.pow(output[0] - sample.target[0], 2);
-    totalLoss += loss;
-
-    const gradOutput = [2 * (output[0] - sample.target[0])];
-    network.BackwardCPU(JSON.stringify([gradOutput]));
-    network.UpdateWeights(JSON.stringify([0.1]));
-  }
-
-  if (epoch % 100 === 0) {
-    console.log(`Epoch ${epoch}: Loss = ${(totalLoss / 4).toFixed(6)}`);
-  }
-}
-
-// Test
-trainingData.forEach((sample) => {
-  const [output] = JSON.parse(
-    network.ForwardCPU(JSON.stringify([sample.input]))
-  );
-  console.log(
-    `${sample.input} → ${output[0].toFixed(4)} (expected ${sample.target[0]})`
-  );
-});
-```
-
-## 🌐 Browser Usage
-
-### Via CDN (UMD)
-
-```html
-<!DOCTYPE html>
-<html>
-  <head>
-    <script src="https://unpkg.com/@openfluke/welvet"></script>
-  </head>
-  <body>
-    <script>
-      (async () => {
-        const { initLoom, ActivationType } = window.Welvet;
-        const loom = await initLoom();
-
-        const network = loom.NewNetwork(4, 1, 1, 1);
-        console.log("LOOM ready!");
-      })();
-    </script>
-  </body>
-</html>
-```
-
-### Via ES Modules
-
-```html
-<!DOCTYPE html>
-<html>
-  <head>
-    <script type="module">
-      import {
-        initLoom,
-        ActivationType,
-      } from "https://unpkg.com/@openfluke/welvet/dist/esm/index.js";
-
-      const loom = await initLoom();
-      const network = loom.NewNetwork(4, 1, 1, 1);
-      console.log("LOOM ready!");
-    </script>
-  </head>
-</html>
-```
-
-## ⚛️ Framework Integration
-
-### React
-
-```tsx
-import { useEffect, useState } from "react";
-import { initLoom, type LoomAPI } from "@openfluke/welvet";
-
-function NeuralNetworkComponent() {
-  const [loom, setLoom] = useState<LoomAPI | null>(null);
-  const [prediction, setPrediction] = useState<number[] | null>(null);
-
-  useEffect(() => {
-    initLoom().then((api) => {
-      setLoom(api);
-
-      // Initialize network
-      const network = api.NewNetwork(4, 1, 1, 2);
-      const layer0 = api.InitDenseLayer(4, 8, 0); // ReLU
-      const layer1 = api.InitDenseLayer(8, 2, 1); // Sigmoid
-
-      network.SetLayer(JSON.stringify([0, 0, 0, JSON.parse(layer0)]));
-      network.SetLayer(JSON.stringify([0, 0, 1, JSON.parse(layer1)]));
-
-      // Make prediction
-      const input = [0.5, 0.3, 0.2, 0.1];
-      const [output] = JSON.parse(network.ForwardCPU(JSON.stringify([input])));
-      setPrediction(output);
-    });
-  }, []);
-
-  if (!loom) return <div>Loading neural network...</div>;
-
-  return (
-    <div>
-      <h2>Prediction: {prediction?.map((v) => v.toFixed(4)).join(", ")}</h2>
-    </div>
-  );
+interface TrainingConfig {
+  Epochs: number;
+  LearningRate: number;
+  LossType?: string;
+  Verbose?: boolean;
+  UseGPU?: boolean;
+  PrintEveryBatch?: number;
+  GradientClip?: number;
 }
 ```
 
-### Vue 3
-
-```vue
-<script setup lang="ts">
-import { ref, onMounted } from "vue";
-import { initLoom, type LoomAPI } from "@openfluke/welvet";
-
-const loom = ref<LoomAPI | null>(null);
-const output = ref<number[] | null>(null);
+## Example
 
-onMounted(async () => {
-  const api = await initLoom();
-  loom.value = api;
+See `example/grid-scatter.ts` for a complete multi-agent training demo:
 
-  const network = api.NewNetwork(2, 1, 1, 1);
-  const layer = api.InitDenseLayer(2, 1, 1); // Sigmoid
-  network.SetLayer(JSON.stringify([0, 0, 0, JSON.parse(layer)]));
-
-  const [result] = JSON.parse(network.ForwardCPU(JSON.stringify([[0.5, 0.5]])));
-  output.value = result;
-});
-</script>
-
-<template>
-  <div v-if="!loom">Loading...</div>
-  <div v-else>
-    <h2>Neural Network Output</h2>
-    <pre>{{ output }}</pre>
-  </div>
-</template>
-```
-
-### Svelte
-
-```svelte
-<script lang="ts">
-  import { onMount } from 'svelte';
-  import { initLoom, type LoomAPI } from '@openfluke/welvet';
-
-  let loom: LoomAPI | null = null;
-  let result: number[] = [];
-
-  onMount(async () => {
-    loom = await initLoom();
-
-    const network = loom.NewNetwork(3, 1, 1, 1);
-    const layer = loom.InitDenseLayer(3, 2, 0); // ReLU
-    network.SetLayer(JSON.stringify([0, 0, 0, JSON.parse(layer)]));
-
-    const [output] = JSON.parse(network.ForwardCPU(JSON.stringify([[1, 2, 3]])));
-    result = output;
-  });
-</script>
-
-{#if !loom}
-  <p>Loading neural network...</p>
-{:else}
-  <h2>Result: {result.join(', ')}</h2>
-{/if}
-```
-
-## 🔧 Advanced Configuration
-
-### Custom WASM Location
-
-```typescript
-const loom = await initLoom({
-  wasmUrl: "/custom/path/loom.wasm",
-});
-```
-
-### Skip Go Runtime Injection
-
-```typescript
-// Useful if you're loading Go runtime separately
-const loom = await initLoom({
-  injectGoRuntime: false,
-});
-```
-
-## 📊 Performance Tips
-
-1. **Batch Processing** - Process multiple inputs together when possible
-2. **Model Caching** - Save trained models to avoid retraining
-3. **Layer Sizing** - Start with smaller layers and scale up as needed
-4. **Learning Rate** - Tune learning rate for faster convergence (typically 0.001 - 0.1)
-5. **Activation Functions** - ReLU often trains faster than Sigmoid/Tanh
-
-## 🐛 Troubleshooting
-
-### WASM fails to load
-
-Ensure your server serves `.wasm` files with the correct MIME type:
-
-```
-Content-Type: application/wasm
+```bash
+cd example
+bun install
+bun run grid-scatter.ts
 ```
 
-### Module not found errors
+Expected output:
 
-Make sure to await the initialization:
-
-```typescript
-const loom = await initLoom(); // Don't forget await!
 ```
-
-### JSON parsing errors
-
-All network methods use JSON string parameters:
-
-```typescript
-// ✅ Correct
-network.ForwardCPU(JSON.stringify([input]));
-
-// ❌ Wrong
-network.ForwardCPU(input);
+🤖 Running Grid Scatter Multi-Agent Training...
+✅ Agent network created!
+Training for 800 epochs with learning rate 0.150
+✅ Training complete!
+Training time: 0.47 seconds
+Initial Loss: 0.252249
+Final Loss: 0.001374
+Improvement: 99.46%
+Total Epochs: 800
 ```
 
-## 🔗 Related Projects
-
-- **Python Package**: [`welvet`](https://pypi.org/project/welvet/) - Python bindings for LOOM
-- **Go Framework**: [LOOM](https://github.com/openfluke/loom) - Original Go implementation
-- **Legacy Package**: [`@openfluke/portal`](https://github.com/openfluke/portal) - Previous generation framework
-
-## 📄 License
+## Layer Types
 
-Apache-2.0 © 2025 OpenFluke
+- `dense` - Fully connected layer
+- `lstm` - Long Short-Term Memory layer
+- `rnn` - Recurrent Neural Network layer
+- `gru` - Gated Recurrent Unit layer
+- `cnn` - Convolutional layer
+- `parallel` - Parallel branches with combine modes:
+  - `add` - Element-wise addition
+  - `concat` - Concatenation
+  - `multiply` - Element-wise multiplication
+  - `grid_scatter` - Multi-agent grid routing
 
-## 🤝 Contributing
+## Activation Functions
 
-Contributions are welcome! Please see the [main repository](https://github.com/openfluke/loom) for guidelines.
+`relu`, `sigmoid`, `tanh`, `softmax`, `gelu`, `swish`, `mish`, `leaky_relu`, `elu`, `selu`
 
-## 📞 Support
+## License
 
-- 🐛 [Report Issues](https://github.com/openfluke/loom/issues)
-- 💬 [Discussions](https://github.com/openfluke/loom/discussions)
-- 📖 [Documentation](https://github.com/openfluke/loom/tree/main/typescript)
+MIT
 
----
+## Links
 
-**Built with ❤️ by the OpenFluke team**
+- [GitHub](https://github.com/openfluke/loom)
+- [WASM Documentation](../wasm/README.md)
+- [Go Examples](../examples/)