// bridge system

useState()→Model Weights·Event Propagation→Forward Pass·Array.map()→Tensor Operation·React diff→Loss Function L = Σ(y−ŷ)²·transition-duration→Learning Rate η·CSS clamp()→σ(x) Activation·Re-render cycle→Training Epoch·Event bubbling→Backpropagation ∂L/∂w·useCallback→Gradient Caching·Promise.all()→Batch Inference·Redux store→Weight Matrix·DevTools profiler→Loss Landscape·useState()→Model Weights·Event Propagation→Forward Pass·Array.map()→Tensor Operation·React diff→Loss Function L = Σ(y−ŷ)²·transition-duration→Learning Rate η·CSS clamp()→σ(x) Activation·Re-render cycle→Training Epoch·Event bubbling→Backpropagation ∂L/∂w·useCallback→Gradient Caching·Promise.all()→Batch Inference·Redux store→Weight Matrix·DevTools profiler→Loss Landscape·

Hassan Al-Rashid
A model that works in your notebook is a hypothesis. A model that works at 3 AM on a solar farm in the Atacama Desert with no internet — that is a product. Start simple. Get something running on the device.

Your first edge model does not need to be complex. In frontend development, you start with basic form validation before building elaborate input pipelines. Edge ML follows the same principle: start with a simple anomaly detector, prove it works on the hardware, then iterate.

Learning Objectives

○Build a simple anomaly detection model using TensorFlow.js
○Understand the structure of sensor data input
○Run inference on a small model within edge constraints
○Measure model size and memory usage

From Validation to Detection

Frontend

Client-Side Validation

function validate(input) {
  return input >= min && input <= max;
}

Machine Learning

Anomaly Detection

function isAnomaly(reading, mean, std) {
  return Math.abs(reading - mean) > 3 * std;
}

Structural Bridge

Client-side form validation checks if input falls within expected bounds. Anomaly detection does the same thing — but it learns the bounds from data instead of hardcoding them.

simple-anomaly-detector.tstypescript

import * as tf from '@tensorflow/tfjs';

// Sensor data: [vibration, temperature, pressure, rpm]
// Each reading is a 4-element vector

// Build a tiny model — must fit in edge device memory
function buildAnomalyDetector(): tf.Sequential {
const model = tf.sequential();

model.add(tf.layers.dense({
  inputShape: [4],
  units: 8,
  activation: 'relu'
}));

model.add(tf.layers.dense({
  units: 1,
  activation: 'sigmoid' // 0 = normal, 1 = anomaly
}));

model.compile({
  optimizer: 'adam',
  loss: 'binaryCrossentropy',
  metrics: ['accuracy']
});

return model;
}

const model = buildAnomalyDetector();

// Check model size — critical for edge deployment
const modelParams = model.countParams();
console.log(`Parameters: ${modelParams}`);  // 49 parameters
console.log(`Estimated size: ${(modelParams * 4 / 1024).toFixed(1)}KB`);  // ~0.2KB

// Run inference on a sensor reading
const sensorReading = tf.tensor2d([[0.85, 72.3, 14.7, 3200]]);
const prediction = model.predict(sensorReading) as tf.Tensor;
const score = prediction.dataSync()[0];
console.log(`Anomaly score: ${score.toFixed(4)}`);

// Clean up tensors — memory management matters on edge devices
sensorReading.dispose();
prediction.dispose();

Notice the dispose() calls. On a device with 512MB RAM, you cannot afford memory leaks. Every tensor must be explicitly freed. This is like removing event listeners in frontend code to prevent memory leaks — except the consequences are a frozen gateway device at a remote site.

Challenge

Build a simple anomaly detector for Terra Grid sensor data.

⚡

Exercise

Beginner~15 min

Build a Simple Anomaly Detector

Build a TensorFlow.js sequential model with: an input dense layer (inputShape [4], 8 units, relu activation) and an output dense layer (1 unit, sigmoid activation). Compile with adam optimizer and binaryCrossentropy loss.

const model = tf.sequential();

// Add a dense layer: inputShape [4], 8 units, relu
// your code here

// Add output layer: 1 unit, sigmoid
// your code here

// Compile with adam and binaryCrossentropy
// your code here

Key Takeaways

✓Start with the simplest model that solves the problem
✓Count parameters to estimate model size: params * 4 bytes for Float32
✓Always dispose tensors on edge devices to prevent memory leaks
✓A 49-parameter model can be a legitimate production anomaly detector

Need a hint?

🧭 Guidance

✅ Solution

Your First Edge Model — Terra Grid | Tensorcraft

Your First Edge Model

Context