<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Raisrujan on Medium]]></title>
        <description><![CDATA[Stories by Raisrujan on Medium]]></description>
        <link>https://medium.com/@raisrujan?source=rss-cbacd1b172ce------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*6uv1tg5yRZtfo474</url>
            <title>Stories by Raisrujan on Medium</title>
            <link>https://medium.com/@raisrujan?source=rss-cbacd1b172ce------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 22:59:05 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@raisrujan/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Death of the Brittle Scraper: How Firecrawl is Solving the Web’s Hardest Data Problems]]></title>
            <link>https://medium.com/@raisrujan/the-death-of-the-brittle-scraper-how-firecrawl-is-solving-the-webs-hardest-data-problems-04b6f70341fa?source=rss-cbacd1b172ce------2</link>
            <guid isPermaLink="false">https://medium.com/p/04b6f70341fa</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[web-scraping-tools]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[firecrawl]]></category>
            <dc:creator><![CDATA[Raisrujan]]></dc:creator>
            <pubDate>Wed, 14 Jan 2026 01:32:43 GMT</pubDate>
            <atom:updated>2026-01-14T03:19:25.926Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yG1HVtsL0zZznevhiAZrUw.png" /></figure><h3>Why modern web architecture broke traditional scraping and how AI-native tools are fixing it.</h3><p>The dream of the internet has always been “data at your fingertips.” But for developers, that data has often felt locked behind a fortress of messy code and defensive walls. For a decade, web scraping was a game of cat-and-mouse. You’d write a script, it would work for 48 hours, and then a minor site update would send your entire pipeline into a tailspin.</p><p>In 2026, the stakes are higher. We don’t just need data; we need <strong>LLM-ready intelligence.</strong> This is where the old tools failed, and where <strong>Firecrawl</strong> has stepped in to redefine the industry.</p><h3>The Three Walls of Web Extraction</h3><p>To understand why Firecrawl is a game-changer, we have to look at why traditional scraping became a nightmare.</p><h4>1. The DOM &amp; Dynamic Frameworks</h4><p>Modern sites are no longer static HTML; they are living applications built on React, Next.js, and Vue.</p><ul><li><strong>The Moving Target:</strong> Frameworks often generate randomized class names (e.g., _container_1abc_xyz). If your scraper relies on these &quot;brittle&quot; CSS selectors, it breaks the moment the site redeploys.</li><li><strong>The Virtual DOM:</strong> Content is often injected by JavaScript seconds after the page loads. Standard scrapers only see a “loading” spinner instead of the actual data.</li></ul><h4>2. The IP Arms Race</h4><p>Websites have become incredibly good at spotting “non-human” behavior. Between <strong>IP rate limiting</strong> and <strong>Sophisticated Fingerprinting</strong> (checking your screen resolution and browser type), a simple script is often blocked by Cloudflare or Akamai before it can even read a single line of text.</p><h4>3. The “Format Hell”</h4><p>Even a successful scrape leaves you with a mountain of “Dirty HTML.” Stripping away &lt;nav&gt; tags, ads, and footers to get clean text for an AI model used to require hours of manual regex or parsing logic.</p><h4>What is Firecrawl?</h4><p>Firecrawl is an AI-native web data platform that abstracts the “plumbing” of web scraping. It treats the entire web like a structured database. Instead of managing proxies or solving CAPTCHAs, you use a simple API to convert any URL into clean, token-efficient Markdown or JSON.</p><h3>How It Fixes the “Old Way”</h3><ul><li><strong>Semantic Extraction:</strong> Firecrawl doesn’t just look for a div; it “understands” the content. You can ask it to “find the product price,” and it uses AI to locate that data regardless of the class name or site structure.</li><li><strong>Built-in Stealth:</strong> It manages <strong>automatic proxy rotation</strong> and anti-bot bypassing out of the box. To a website, Firecrawl looks like a regular human browsing from a laptop.</li><li><strong>Markdown-First Design:</strong> Firecrawl converts the messy web into Markdown the “native language” of LLMs. This reduces data size by up to 90%, making your AI agents faster and significantly cheaper to run.</li></ul><h3>The Technical Edge: A Tool for Every Task</h3><p>Firecrawl provides a suite of endpoints that handle the entire lifecycle of data discovery:</p><ul><li><strong>/map</strong>: Quickly scans a domain to find all sub-URLs without full scraping.</li><li><strong>/scrape</strong>: Converts a single page into perfect Markdown.</li><li><strong>/crawl</strong>: Recursively traverses a site, mapping and scraping every page automatically.</li><li><strong>/extract</strong>: Provide a JSON schema, and Firecrawl uses AI to pull specific data points (like job titles or stock prices) with 100% accuracy.</li></ul><h3>Why Firecrawl is “RAG-Ready”</h3><p>If you are building <strong>Retrieval-Augmented Generation (RAG)</strong> systems, Firecrawl is your most valuable asset. By providing clean, structured data with preserved metadata (URLs and titles), it allows your AI to provide accurate citations and avoid hallucinations. It doesn’t just give your AI data; it gives it <strong>context.</strong></p><h3>Conclusion: The New Standard</h3><p>The bottleneck for AI today isn’t the model it’s the data. We shouldn’t be spending our engineering hours fixing broken CSS selectors.</p><p>Firecrawl has effectively turned the messy, unpredictable web into a clean API. Whether you are building a competitive intelligence engine or a real-time AI researcher, Firecrawl allows you to focus on <strong>what you do with the data</strong>, rather than how to get it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=04b6f70341fa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Running Machine Learning on ESP32: Bringing AI to Edge Computing]]></title>
            <link>https://medium.com/@raisrujan/running-machine-learning-on-esp32-bringing-ai-to-edge-computingesp32-is-an-edge-device-4349b79f848a?source=rss-cbacd1b172ce------2</link>
            <guid isPermaLink="false">https://medium.com/p/4349b79f848a</guid>
            <category><![CDATA[tinyml]]></category>
            <category><![CDATA[esp32]]></category>
            <category><![CDATA[iot]]></category>
            <category><![CDATA[machine-learning-ai]]></category>
            <category><![CDATA[edge-computing]]></category>
            <dc:creator><![CDATA[Raisrujan]]></dc:creator>
            <pubDate>Sun, 16 Mar 2025 08:17:52 GMT</pubDate>
            <atom:updated>2025-03-16T08:19:44.996Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*CNlsvwc1AFPIisPp" /></figure><h3>Introduction</h3><p>The <strong>ESP32 microcontroller</strong> has become a popular choice for IoT applications due to its affordability, low power consumption, and WiFi/Bluetooth capabilities. But did you know it’s also powerful enough to run <strong>machine learning (ML) models</strong>? With the rise of <strong>TinyML (tiny machine learning)</strong>, deploying ML models on resource-constrained devices like the ESP32 is now a reality.</p><p>This guide will walk you through how to run ML models on the ESP32, covering:</p><ul><li>Why ESP32 is a great choice for ML</li><li>How to set up ML on ESP32</li><li>Training and deploying a model using <strong>TensorFlow Lite for Microcontrollers</strong></li><li>Running a real-world <strong>classification model</strong></li><li>Optimizing ML models for the ESP32</li><li>Challenges and best practices</li></ul><p>By the end, you’ll have a working knowledge of <strong>edge AI</strong> and how to implement ML on the ESP32 for smart IoT applications.</p><h3>Why Run ML on ESP32?</h3><p>The ESP32 is designed primarily as an IoT microcontroller, not a machine learning powerhouse. However, with <strong>TinyML</strong>, we can run lightweight ML models directly on the device, enabling real-time, low-latency decision-making. Here’s why it’s a game-changer:</p><p><strong>Low Cost &amp; Accessibility:</strong> ESP32 boards are cheap (~$5-$10) and widely available.<br> <strong>Power Efficiency:</strong> Ideal for battery-powered devices in remote locations.<br> <strong>Connectivity:</strong> Built-in WiFi/Bluetooth for cloud integration.<br> <strong>Edge AI Potential:</strong> Reduces reliance on cloud computing, improving speed and privacy.</p><h3>Setting Up ML on ESP32</h3><h3>1. Required Hardware &amp; Software</h3><p>To get started, you’ll need:</p><p><strong>Hardware:</strong></p><ul><li>ESP32 Dev Board (ESP32-S3 recommended for ML)</li><li>Sensors (e.g., temperature, accelerometer, microphone) for data collection</li><li>USB cable and power source</li></ul><p><strong>Software:</strong></p><ul><li>Arduino IDE or ESP-IDF</li><li>TensorFlow Lite for Microcontrollers (TFLM)</li><li>Python for data preprocessing</li><li>Edge Impulse (optional, for model training)</li></ul><h3>Installing TensorFlow Lite for Microcontrollers</h3><p>TensorFlow Lite for Microcontrollers (TFLM) is a lightweight version of TensorFlow designed for small devices. Here’s how to install it:</p><ol><li><strong>Install the Arduino IDE &amp; ESP32 Board Support Package</strong></li><li><strong>Add the TensorFlow Lite library</strong> in the Arduino Library Manager.</li><li><strong>Set up the development environment</strong> (Arduino or PlatformIO).</li></ol><p>Now, you’re ready to train and deploy a model!</p><h3>Training &amp; Deploying a Model on ESP32</h3><h3>Step 1: Choosing a Model</h3><p>Since ESP32 has limited resources (typically 520 KB SRAM, 4MB Flash), you need an efficient ML model. Some good options include:</p><ul><li><strong>Classification models:</strong> Predict whether a sensor reading indicates an event.</li><li><strong>Regression models:</strong> Predict numerical values, such as temperature trends.</li><li><strong>Anomaly detection:</strong> Detect abnormal behavior in devices.</li></ul><p>For this example, let’s train a <strong>logistic regression model</strong> to classify flowers from the <strong>Iris dataset</strong> (a simple ML problem).</p><h3>Step 2: Training the Model</h3><p>Training happens <strong>off the ESP32</strong> on a computer. We use Python and TensorFlow:</p><pre><br>import tensorflow as tf<br>from sklearn.datasets import load_iris<br>from sklearn.model_selection import train_test_split<br>from sklearn.preprocessing import StandardScaler<br><br># Load dataset<br>iris = load_iris()<br>X = iris.data<br>y = iris.target<br><br># Preprocess data<br>scaler = StandardScaler()<br>X_scaled = scaler.fit_transform(X)<br><br># Train a simple model<br>model = tf.keras.Sequential([<br>    tf.keras.layers.Dense(10, activation=&#39;relu&#39;),<br>    tf.keras.layers.Dense(3, activation=&#39;softmax&#39;)<br>])<br><br>model.compile(optimizer=&#39;adam&#39;, loss=&#39;sparse_categorical_crossentropy&#39;, metrics=[&#39;accuracy&#39;])<br>model.fit(X_scaled, y, epochs=50)</pre><p>Once trained, we <strong>convert the model to TensorFlow Lite format</strong> to run on ESP32:</p><pre>converter = tf.lite.TFLiteConverter.from_keras_model(model)<br>tflite_model = converter.convert()<br># Save the model<br>with open(&quot;iris_model.tflite&quot;, &quot;wb&quot;) as f:<br>    f.write(tflite_model)</pre><h3>Step 3: Deploying the Model to ESP32</h3><p>Now, we need to upload the model to the ESP32. We do this using <strong>Arduino IDE</strong>:</p><ol><li>Convert the .tflite file to a <strong>C array</strong> using a converter tool:</li></ol><pre>xxd -i iris_model.tflite &gt; iris_model.h</pre><ul><li>Add the iris_model.h file to the ESP32 project.</li></ul><p>2. Write an Arduino script to load and run the model:</p><pre>#include &quot;iris_model.h&quot;<br>#include &lt;tensorflow/lite/micro/all_ops_resolver.h&gt;<br>#include &lt;tensorflow/lite/micro/micro_interpreter.h&gt;<br>#include &lt;tensorflow/lite/schema/schema_generated.h&gt;<br><br>const tflite::Model* model = tflite::GetModel(iris_model);<br>tflite::MicroInterpreter interpreter(model, ...);<br>interpreter.AllocateTensors();<br><br>// Add input data<br>float input_data[] = {5.1, 3.5, 1.4, 0.2}; // Example Iris data<br>float* output_data = interpreter.output(0)-&gt;data.f;<br><br>// Run inference<br>interpreter.Invoke();<br>Serial.println(output_data[0]); // Output prediction</pre><h3>Optimizing ML Models for ESP32</h3><p>Since ESP32 has limited RAM and processing power, <strong>optimizations are crucial</strong>:</p><p><strong>Use Model Quantization</strong></p><ul><li>Converts float32 to int8, reducing model size by <strong>75%</strong>.</li><li>Example:</li></ul><pre>converter.optimizations = [tf.lite.Optimize.DEFAULT]</pre><p><strong>Reduce Model Complexity</strong></p><ul><li>Fewer layers and neurons = less RAM usage.</li><li>Avoid deep neural networks (DNNs) on ESP32.</li></ul><p><strong>Use ESP32-S3 for Better Performance</strong></p><ul><li><strong>ESP32-S3 has vector instructions (AI acceleration).</strong></li><li>Handles larger models efficiently.</li></ul><h3>Real-World Use Cases for ML on ESP32</h3><p>Now that we can run ML models on ESP32, here are some <strong>practical applications</strong>:</p><ul><li><strong>Smart Agriculture:</strong> Predict soil moisture levels to optimize irrigation.</li><li><strong>Home Automation:</strong> Recognize voice commands for controlling lights.</li><li><strong>Predictive Maintenance:</strong> Detect anomalies in industrial machines.</li><li><strong>Gesture Recognition:</strong> Classify hand movements using an accelerometer.</li><li><strong>Wildlife Monitoring:</strong> Identify animal sounds in forests.</li></ul><p>These applications demonstrate the <strong>power of TinyML</strong> on ESP32!</p><h3>Challenges &amp; Best Practices</h3><p>Running ML on ESP32 comes with some <strong>challenges</strong>:</p><p><strong>Limited RAM &amp; Processing Power:</strong><br> Use <strong>quantization</strong> and <strong>simpler models</strong>.</p><p><strong>No GPU or TPU for AI acceleration:</strong><br> Use <strong>ESP32-S3</strong> for better ML support.</p><p><strong>Difficult Debugging &amp; Deployment:</strong><br> Use <strong>Edge Impulse</strong> for streamlined ML workflows.</p><h3>Conclusion: TinyML + ESP32 = Future of Edge AI</h3><p>With <strong>TinyML</strong>, the ESP32 is no longer just an IoT device — it’s a <strong>powerful edge AI platform</strong>. We can build intelligent, real-time ML applications without relying on the cloud by leveraging <strong>TensorFlow Lite</strong>, efficient models, and the ESP32-S3’s hardware acceleration.</p><p>Whether you’re working on <strong>smart homes, industrial IoT, or real-time monitoring</strong>, <strong>ML on ESP32</strong> opens the door to <strong>a new era of smart devices</strong>. Start experimenting today, and bring AI to the edge!</p><p>Did you find this guide helpful? Let me know in the comments, and share your own ML projects running on ESP32!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4349b79f848a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>