<?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 aravinth C on Medium]]></title>
        <description><![CDATA[Stories by aravinth C on Medium]]></description>
        <link>https://medium.com/@aravinthc18?source=rss-87e95b073259------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*vEA98pwfWFOsTeCAfcz5yw.jpeg</url>
            <title>Stories by aravinth C on Medium</title>
            <link>https://medium.com/@aravinthc18?source=rss-87e95b073259------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 19:17:50 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aravinthc18/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[Making OpenCode Look the Same Everywhere: My Journey Toward Exact TUI and Desktop Theme Alignment]]></title>
            <link>https://medium.com/@aravinthc18/making-opencode-look-the-same-everywhere-my-journey-toward-exact-tui-and-desktop-theme-alignment-d6419a0131d0?source=rss-87e95b073259------2</link>
            <guid isPermaLink="false">https://medium.com/p/d6419a0131d0</guid>
            <category><![CDATA[opencode-theme]]></category>
            <category><![CDATA[open-code]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[themes]]></category>
            <category><![CDATA[desktop]]></category>
            <dc:creator><![CDATA[aravinth C]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 05:54:49 GMT</pubDate>
            <atom:updated>2026-11-16T05:54:49.408Z</atom:updated>
            <content:encoded><![CDATA[<h3><em>Bringing consistency to OpenCode’s terminal and desktop experience through a custom theme.</em></h3><h3>Introduction</h3><p>One thing I appreciate about OpenCode is its simplicity, developer-focused experience, and clean interface. Whether working in the terminal or using the desktop application, OpenCode provides a productive environment for AI-assisted development.</p><p>However, while exploring its theming capabilities, I noticed something interesting.</p><p><strong>The terminal TUI and desktop application did not share exactly the same colour palette — even when both were configured to use the OpenCode theme.</strong></p><p>This small visual difference led me to investigate how OpenCode handles themes internally.</p><h3>The Problem: Two Theme Systems, One Product</h3><p>OpenCode currently uses different theme systems for its terminal and desktop interfaces.</p><p>The terminal TUI uses a theme.json schema based on colour ramps and semantic theme tokens. Custom themes can be loaded from:</p><pre>&lt;configDir&gt;/themes/*.json</pre><p>The desktop application, on the other hand, uses a separate desktop-theme.json schema with a flat colour palette and overrides. Its themes are registered through a hardcoded import map in the application bundle.</p><p>As a result, selecting the default OpenCode theme in both applications does not guarantee an identical appearance.</p><p>The conversation area may look similar, but the desktop window chrome — including the title bar and tab strip — can use different colours.</p><p>For developers who care about visual consistency, this becomes noticeable.</p><h3>Investigating the Original OpenCode Palette</h3><p>Instead of manually approximating the colours, I wanted to reproduce the original terminal palette as accurately as possible.</p><p>I extracted the built-in opencode theme from the OpenCode CLI binary, specifically version 2.0.9.</p><p>The objective was simple:</p><p><strong>Preserve the original terminal colours without introducing any visual approximation.</strong></p><p>I then mapped the same colour values into the desktop theme schema.</p><p>This approach allowed me to maintain the original palette while adapting it to two different theme engines.</p><h3>My Solution: OpenCode Exact Theme</h3><p>I created two complementary theme files.</p><h3>1. opencode-exact.json</h3><p>This file follows the official terminal theme.json schema.</p><p>It reproduces the built-in OpenCode theme and can be installed in the user’s themes directory:</p><pre>.config\opencode\themes\</pre><p>The theme registers under the ID:</p><pre>opencode-exact</pre><p>It allows users to pin the exact terminal colour palette rather than depending on the theme bundled with a particular CLI version.</p><h3>2. opencode-terminal.desktop-theme.json</h3><p>The second file maps the same colours into the desktop application’s desktop-theme.json structure.</p><p>It includes the palette and overrides required to represent the terminal colours in the desktop interface.</p><p>This file can serve as a reference implementation or as a starting point for integrating the theme into the desktop application’s import map.</p><h3>3. README Documentation</h3><p>I also prepared a README containing installation instructions, palette information, and submission notes.</p><p>The goal was to make the theme easy for other developers to understand, install, and review.</p><h3>The Colour Palette</h3><p>The theme preserves the original OpenCode dark palette.</p><p>Theme TokenColourBackground Base#0a0a0aBackground Raised Base#141414Text Base#eeeeeePrimary / Interactive#fab283Success#7fd88fWarning#f5a742Error#e06c75Info#56b6c2</p><p>These colours provide the familiar OpenCode appearance while maintaining consistency between the terminal and desktop representations.</p><h3>Validation and Accuracy</h3><p>I also validated the theme serialization process.</p><p>The theme successfully passed round-trip verification:</p><pre>Parse → Stringify → Parse</pre><p>The resulting structure remains identical, ensuring that serialization does not introduce unexpected changes to the theme data.</p><p>This was important because the goal was not simply to create a visually similar theme, but to preserve the exact colour definitions.</p><h3>Why This Matters</h3><p>A theme is more than a collection of colours.</p><p>For developers who frequently switch between terminal and desktop environments, visual consistency creates a more unified experience.</p><p>There is also a practical benefit to maintaining a pinned theme.</p><p>When the built-in theme changes during an OpenCode upgrade, users may experience unexpected colour differences. A pinned theme provides a stable alternative for those who prefer the original appearance.</p><h3>My Contribution to the OpenCode Community</h3><p>I would like to propose this theme for inclusion as a built-in option in a future OpenCode release.</p><p>My intention is to contribute something useful to the community rather than keeping the solution limited to my local environment.</p><p>The implementation includes:</p><ul><li>Exact terminal palette extraction.</li><li>Desktop palette adaptation.</li><li>Theme configuration files.</li><li>Installation documentation.</li><li>Serialization validation.</li></ul><p>I am also interested in understanding the maintainers’ preferred approach for integrating custom desktop themes into OpenCode.</p><h3>What’s Next?</h3><p>I plan to share this work with the OpenCode community through GitHub and seek feedback from maintainers.</p><p>If the approach is accepted, I would be happy to contribute the necessary changes through a pull request.</p><p>The broader idea is simple:</p><p><strong>One OpenCode theme. One consistent colour experience. Across terminal and desktop.</strong></p><h3>Final Thoughts</h3><p>This project started with a small observation: the terminal and desktop versions of OpenCode did not look exactly the same.</p><p>By investigating the theme architecture, extracting the original palette, and mapping it across both theme systems, I was able to build a solution that preserves the original OpenCode colours.</p><p>Sometimes, the most interesting developer contributions begin with noticing a small inconsistency and deciding to investigate it.</p><p>That is exactly how this project started.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pMeT7galJFe_KdhUfL6HGg.png" /></figure><p><em>Project: OpenCode Exact Theme</em><br><em>Author: Aravinth Chinnasamy</em></p><p>GitHub: <a href="https://github.com/anomalyco/opencode/issues/64163">https://github.com/anomalyco/opencode/issues/64163</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JBihw1jTP2tYkWgzlWo3Ow.png" /></figure><p>#OpenCode #OpenSource #DeveloperTools #Terminal #Theming #AI #OpenSourceContribution #SoftwareEngineering</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d6419a0131d0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ Installing and Configuring Neo4j on an AWS Linux Server]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@aravinthc18/installing-and-configuring-neo4j-on-an-aws-linux-server-2f65f7ae087c?source=rss-87e95b073259------2"><img src="https://cdn-images-1.medium.com/max/1525/1*7QJVDANa24SzYZhMDyr5dA.png" width="1525"></a></p><p class="medium-feed-snippet">&#x201C;A graph becomes powerful when your data is connected, not just stored.&#x201D;</p><p class="medium-feed-link"><a href="https://medium.com/@aravinthc18/installing-and-configuring-neo4j-on-an-aws-linux-server-2f65f7ae087c?source=rss-87e95b073259------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@aravinthc18/installing-and-configuring-neo4j-on-an-aws-linux-server-2f65f7ae087c?source=rss-87e95b073259------2</link>
            <guid isPermaLink="false">https://medium.com/p/2f65f7ae087c</guid>
            <category><![CDATA[fastapi]]></category>
            <category><![CDATA[aws-s3]]></category>
            <category><![CDATA[neo4j]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[neo4j-graph-database]]></category>
            <dc:creator><![CDATA[aravinth C]]></dc:creator>
            <pubDate>Sat, 19 Sep 2026 07:58:27 GMT</pubDate>
            <atom:updated>2026-09-19T07:58:27.927Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[ UV: A Faster and Simpler Way to Manage Python Projects]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">&#x201C;Python development should focus on building applications &#x2014; not fighting with environments and dependencies.&#x201D;</p><p class="medium-feed-link"><a href="https://medium.com/@aravinthc18/uv-a-faster-and-simpler-way-to-manage-python-projects-539e9e07167c?source=rss-87e95b073259------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@aravinthc18/uv-a-faster-and-simpler-way-to-manage-python-projects-539e9e07167c?source=rss-87e95b073259------2</link>
            <guid isPermaLink="false">https://medium.com/p/539e9e07167c</guid>
            <category><![CDATA[uvx]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[uv]]></category>
            <category><![CDATA[pip-install]]></category>
            <category><![CDATA[python]]></category>
            <dc:creator><![CDATA[aravinth C]]></dc:creator>
            <pubDate>Mon, 07 Sep 2026 15:55:45 GMT</pubDate>
            <atom:updated>2026-09-07T15:55:45.373Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[ Setting Up MongoDB Locally on Windows: Server, Shell & Compass]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@aravinthc18/setting-up-mongodb-locally-on-windows-server-shell-compass-4b24acbe45f5?source=rss-87e95b073259------2"><img src="https://cdn-images-1.medium.com/max/1664/1*cf0onRwI640nb-7BsnrzDA.png" width="1664"></a></p><p class="medium-feed-snippet">&#x201C;A smooth MongoDB setup starts with installing the components in the right order.&#x201D;</p><p class="medium-feed-link"><a href="https://medium.com/@aravinthc18/setting-up-mongodb-locally-on-windows-server-shell-compass-4b24acbe45f5?source=rss-87e95b073259------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@aravinthc18/setting-up-mongodb-locally-on-windows-server-shell-compass-4b24acbe45f5?source=rss-87e95b073259------2</link>
            <guid isPermaLink="false">https://medium.com/p/4b24acbe45f5</guid>
            <category><![CDATA[mongodb-atlas]]></category>
            <category><![CDATA[python-programming]]></category>
            <category><![CDATA[mongodb-compass]]></category>
            <category><![CDATA[mongodb]]></category>
            <dc:creator><![CDATA[aravinth C]]></dc:creator>
            <pubDate>Sun, 30 Aug 2026 04:47:22 GMT</pubDate>
            <atom:updated>2026-08-30T04:47:22.517Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Building a High-Performance RAG Architecture with ColBERTv2 & Flask ⚡]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">A step-by-step guide to implementing late-interaction semantic search using RAGatouille, LangChain, and FAISS.</p><p class="medium-feed-link"><a href="https://medium.com/@aravinthc18/building-a-high-performance-rag-architecture-with-colbertv2-flask-8c5e36ad16e6?source=rss-87e95b073259------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@aravinthc18/building-a-high-performance-rag-architecture-with-colbertv2-flask-8c5e36ad16e6?source=rss-87e95b073259------2</link>
            <guid isPermaLink="false">https://medium.com/p/8c5e36ad16e6</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[machine-learning-ai]]></category>
            <category><![CDATA[retrieval-augmented-gen]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[aravinth C]]></dc:creator>
            <pubDate>Wed, 19 Aug 2026 10:07:14 GMT</pubDate>
            <atom:updated>2026-08-19T10:07:14.956Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[ Running RAGatouille with ColBERT v2.0 on Windows: Problems I Faced and How I Fixed Them]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">If you are building a RAG (Retrieval-Augmented Generation) application, you have probably heard of ColBERT.</p><p class="medium-feed-link"><a href="https://medium.com/@aravinthc18/running-ragatouille-with-colbert-v2-0-on-windows-problems-i-faced-and-how-i-fixed-them-0a7eda82d05d?source=rss-87e95b073259------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@aravinthc18/running-ragatouille-with-colbert-v2-0-on-windows-problems-i-faced-and-how-i-fixed-them-0a7eda82d05d?source=rss-87e95b073259------2</link>
            <guid isPermaLink="false">https://medium.com/p/0a7eda82d05d</guid>
            <category><![CDATA[faiss]]></category>
            <category><![CDATA[cuda]]></category>
            <category><![CDATA[ratatouille]]></category>
            <category><![CDATA[retrieval-augmented-gen]]></category>
            <category><![CDATA[colbert]]></category>
            <dc:creator><![CDATA[aravinth C]]></dc:creator>
            <pubDate>Tue, 11 Aug 2026 04:06:32 GMT</pubDate>
            <atom:updated>2026-08-11T04:06:32.117Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[ Deploying Your MCP Server on Render and Connecting It to Claude AI]]></title>
            <link>https://medium.com/@aravinthc18/deploying-your-mcp-server-on-render-and-connecting-it-to-claude-ai-16b8ab262ce7?source=rss-87e95b073259------2</link>
            <guid isPermaLink="false">https://medium.com/p/16b8ab262ce7</guid>
            <category><![CDATA[claude-ai]]></category>
            <category><![CDATA[render-site]]></category>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[https]]></category>
            <dc:creator><![CDATA[aravinth C]]></dc:creator>
            <pubDate>Tue, 28 Jul 2026 12:38:58 GMT</pubDate>
            <atom:updated>2026-08-02T06:44:41.910Z</atom:updated>
            <content:encoded><![CDATA[<h3>🌍 Taking Our MCP Server to the Next Level</h3><p>In the previous article, <a href="https://medium.com/@aravinthc18/building-your-first-mcp-server-in-python-using-fastmcp-ca3ccf98859c">🚀Building Your First MCP Server in Python Using FastMCP</a> , we built a basic MCP server using FastMCP and tested it locally with the MCP Inspector. That was a great way to understand the fundamentals of creating and exposing tools through the Model Context Protocol (MCP).</p><p>Now, let’s take the next step by deploying the MCP server to <strong>Render</strong>, making it accessible over <strong>HTTPS</strong>, and connecting it directly to <strong>Claude AI</strong> as a <strong>Custom Connector</strong>. Once connected, Claude can automatically discover and invoke the tools exposed by your MCP server, allowing you to integrate your own Python applications into real AI workflows.</p><p>In this tutorial, we’ll deploy the server on <strong>Render</strong> and connect it as a <strong>Custom Connector</strong> in Claude AI.</p><h3>📝 Updated MCP Server</h3><pre>from mcp.server.fastmcp import FastMCP<br><br>from mcp.server.transport_security import TransportSecuritySettings<br><br>mcp = FastMCP(<br>    &quot;calculator_server&quot;,<br>    json_response=True,<br>    host=&quot;0.0.0.0&quot;,<br>    transport_security=TransportSecuritySettings(<br>        enable_dns_rebinding_protection=True,<br>        allowed_hosts=[&quot;your-server-name.onrender.com&quot;, &quot;your-server-name.onrender.com:*&quot;],<br>        allowed_origins=[&quot;https://your-server-name.onrender.com&quot;],<br>    ),<br>)<br><br><br>@mcp.tool()<br>def add(a: int, b: int):<br>    print(&quot;Calling add&quot;)<br>    return a + b<br><br><br>@mcp.tool()<br>def sub(a: int, b: int):<br>    print(&quot;Calling sub&quot;)<br>    return a - b<br><br><br>@mcp.tool()<br>def divide(a: int, b: int):<br>    print(&quot;Calling div&quot;)<br>    return a // b<br><br><br>@mcp.tool()<br>def mul(a: int, b: int):<br>    print(&quot;Calling mul&quot;)<br>    return a * b<br><br><br>transport_security = TransportSecuritySettings(enable_dns_rebinding_protection=False)<br>app = mcp.streamable_http_app()</pre><h3>🔒 Why Disable TransportSecuritySettings?</h3><p>When running an MCP server locally, DNS rebinding protection helps prevent malicious websites from accessing your local services.</p><p>However, when your application is deployed behind a cloud platform like <strong>Render</strong>, the requests are already coming through a public HTTPS domain managed by the hosting provider.</p><p>For this deployment, we disable DNS rebinding protection:</p><pre>transport_security=TransportSecuritySettings(<br>    enable_dns_rebinding_protection=False<br>)</pre><p>This allows Claude AI to communicate with your hosted MCP server without DNS rebinding restrictions.</p><blockquote><strong><em>Note:</em></strong><em> Only disable this when your server is intentionally exposed through a trusted HTTPS hosting provider such as Render.</em></blockquote><h3>🚀 Deploy the Server on Render</h3><p>Deploy your Python application to Render.</p><p>Once deployment finishes, you’ll receive a public HTTPS URL similar to:</p><pre>https://your-server-name.onrender.com</pre><p>Your MCP endpoint will be:</p><pre>https://your-server-name.onrender.com/mcp</pre><p>Verify the endpoint is reachable before connecting it to Claude.</p><h3>🤖 Connect the MCP Server to Claude AI</h3><p>Once your server is deployed, follow these steps.</p><h3>Step 1</h3><p>Open <strong>Claude AI</strong>.</p><h3>Step 2</h3><p>Click your <strong>Profile</strong> in the lower-left corner.</p><h3>Step 3</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/277/1*_hYfsDrwC9jhq9ybSvkzuQ.png" /></figure><p>Select <strong>Settings</strong> (or <strong>Customize Claude</strong>, depending on the current interface).</p><h3>Step 4</h3><p>Open the <strong>Connectors</strong> section.</p><h3>Step 5</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/933/1*d70k4Z5husJLFV8TKOd85g.png" /></figure><p>Click <strong>Add Connector</strong> in the top-right corner.</p><h3>Step 6</h3><p>Choose <strong>Add Custom Connector</strong>.</p><h3>Step 7</h3><p>Enter the following details:</p><p><strong>Server Name</strong></p><pre>calculator_server</pre><p>Use the same name you specified in your Python code:</p><pre>mcp = FastMCP(&quot;calculator_server&quot;)</pre><p><strong>Server URL</strong></p><pre>https://your-server-name.onrender.com/mcp</pre><p>Replace it with your own Render deployment URL.</p><h3>Step 8</h3><p>Claude will begin connecting to your MCP server.</p><p>You should see the connection process complete successfully.</p><h3>Step 9</h3><p>Claude automatically discovers the tools exposed by your server.</p><p>For this example, it will find:</p><ul><li>➕ add</li><li>➖ sub</li><li>✖️ mul</li><li>➗ divide</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/791/1*ap-P77Uu003wrowCkPqLIQ.png" /></figure><p>No additional configuration is required.</p><h3>💬 Start Using Your MCP Server</h3><p>Open a <strong>New Chat</strong> in Claude AI.</p><p>To encourage Claude to use your connector, you can begin with a prompt such as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/952/1*SEXrGTrfCn_m2kBxYdLHvg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HLXDEMZIOuy5OwoVZAPZOA.png" /></figure><blockquote><strong><em>Use my MCP server whenever one of its tools can help answer my requests in this conversation.</em></strong></blockquote><p>Now try asking questions like:</p><blockquote><em>What is 245 + 678?</em></blockquote><p>Claude can invoke the <strong>add</strong> tool from your MCP server.</p><p>Or:</p><blockquote><em>Multiply 145 by 76.</em></blockquote><p>Claude can call the <strong>mul</strong> tool automatically.</p><p>Although this example uses simple calculator functions, the same approach works for much more powerful tools, such as:</p><ul><li>📄 Reading PDFs</li><li>🗄️ Database queries</li><li>📧 Sending emails</li><li>📊 Generating reports</li><li>🤖 Calling AI models</li><li>📁 File management</li><li>🌐 External API integrations</li></ul><h3>🎉 Congratulations!</h3><p>You have successfully:</p><p>✅ Built an MCP Server using FastMCP</p><p>✅ Hosted it publicly on Render</p><p>✅ Connected it to Claude AI as a Custom Connector</p><p>✅ Allowed Claude to discover and invoke your tools automatically</p><p>This simple calculator project is just the beginning. By replacing these basic functions with your own business logic, you can enable Claude to interact securely with your applications, databases, APIs, and AI services through the Model Context Protocol.</p><p><strong>One small correction for the article:</strong> avoid telling readers to use a prompt like <strong>“Use my MCP server for all chat process and do whatever you need to build or call.”</strong> Claude decides when to use available tools based on the user’s request. A better instruction is:</p><blockquote><strong><em>“Use my MCP server whenever one of its tools is relevant to answering my requests in this conversation.”</em></strong></blockquote><p>This more accurately reflects how MCP tool invocation works.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=16b8ab262ce7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building Your First MCP Server in Python Using FastMCP]]></title>
            <link>https://medium.com/@aravinthc18/building-your-first-mcp-server-in-python-using-fastmcp-ca3ccf98859c?source=rss-87e95b073259------2</link>
            <guid isPermaLink="false">https://medium.com/p/ca3ccf98859c</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[mcp-server]]></category>
            <dc:creator><![CDATA[aravinth C]]></dc:creator>
            <pubDate>Tue, 21 Jul 2026 12:46:44 GMT</pubDate>
            <atom:updated>2026-07-21T12:46:44.600Z</atom:updated>
            <content:encoded><![CDATA[<p>The Model Context Protocol (MCP) is an open protocol that enables AI assistants to communicate with external tools and services. Instead of embedding business logic inside an AI application, MCP allows you to expose your functions as reusable tools.</p><p>In this tutorial, we’ll build a simple MCP server using <strong>FastMCP</strong> that provides four calculator operations: addition, subtraction, multiplication, and division.</p><p>calculator operations: addition, subtraction, multiplication, and division.</p><h3><strong>⚙️</strong>Prerequisites</h3><p>Before getting started, make sure you have:</p><ul><li>Python 3.10 or later</li><li>pip installed</li><li>Basic knowledge of Python</li></ul><p>📥Install the MCP package:</p><pre>pip install &quot;mcp[cli]&quot;</pre><h3>🔑Creating the MCP Server</h3><p>Create a file named server.py.</p><pre>from mcp.server.fastmcp import FastMCP<br>mcp = FastMCP(&quot;Test Server&quot;, json_response=True)<br><br>@mcp.tool()<br>def add(a: int, b: int):<br>    return a + b<br><br>@mcp.tool()<br>def sub(a: int, b: int):<br>    return a - b<br><br>@mcp.tool()<br>def mul(a: int, b: int):<br>    return a * b<br><br>@mcp.tool()<br>def divide(a: int, b: int):<br>    return a / b<br><br>if __name__ == &quot;__main__&quot;:<br>    mcp.run(transport=&quot;streamable-http&quot;)</pre><h3>🔍Understanding the Code</h3><h3>Import FastMCP</h3><pre>from mcp.server.fastmcp import FastMCP</pre><p>This imports the FastMCP framework used to create an MCP server.</p><h3>🌐Create the Server</h3><pre>mcp = FastMCP(&quot;Test Server&quot;, json_response=True)</pre><ul><li><strong>Test Server</strong> is the server name.</li><li><strong>json_response=True</strong> returns responses in JSON format, making integration easier.</li></ul><h3>✅Register Tools</h3><p>Each function becomes an MCP tool by decorating it with @mcp.tool().</p><h3>Addition</h3><pre>@mcp.tool()<br>def add(a: int, b: int):<br>    return a + b</pre><h3>Subtraction</h3><pre>@mcp.tool()<br>def sub(a: int, b: int):<br>    return a - b</pre><h3>Multiplication</h3><pre>@mcp.tool()<br>def mul(a: int, b: int):<br>    return a * b</pre><h3>Division</h3><pre>@mcp.tool()<br>def divide(a: int, b: int):<br>    return a / b</pre><h3>▶️ Run the MCP Server</h3><p>Start your MCP server by executing:</p><pre>python server.py</pre><p>If everything starts successfully, you’ll see output similar to this:</p><pre>INFO:     Started server process [18728]<br>INFO:     Waiting for application startup.<br>INFO:     StreamableHTTP session manager started<br>INFO:     Application startup complete.<br>INFO:     Uvicorn running on http://127.0.0.1:8000 (Press CTRL+C to quit)</pre><p>✅ Your MCP server is now running at:</p><pre>http://127.0.0.1:8000</pre><h3>🛠️ Launch the MCP Inspector</h3><p>Open <strong>another Command Prompt</strong> or <strong>Terminal</strong> (keep the server running) and execute:</p><pre>npx @modelcontextprotocol/inspector</pre><p>After a few seconds, you’ll see output similar to:</p><pre>Starting MCP inspector...</pre><pre>⚙️ Proxy server listening on localhost:6277</pre><pre>🔑 Session token:<br>TOKEN</pre><pre>Use this token to authenticate requests or set<br>DANGEROUSLY_OMIT_AUTH=true to disable authentication.</pre><pre>🚀 MCP Inspector is up and running at:</pre><pre>http://localhost:6274/?MCP_PROXY_AUTH_TOKEN=TOKEN</pre><pre>🌐 Opening browser...</pre><h3>📌 What Do These Messages Mean?</h3><p>IconMeaning▶️Starts your FastMCP server⚙️MCP Inspector proxy has started🔑Authentication token for connecting to the proxy🚀Inspector web interface is ready🌐Browser is automatically opening</p><h3>🌍 Open the Inspector</h3><p>If the browser doesn’t open automatically, copy and paste the generated URL into your browser:</p><pre>http://localhost:6274/?MCP_PROXY_AUTH_TOKEN=TOKEN</pre><p>You’ll see the <strong>MCP Inspector</strong> interface, where you can:</p><ul><li>🔍 Discover available tools</li><li>▶️ Execute tools interactively</li><li>📥 Send JSON parameters</li><li>📤 View responses</li><li>🐞 Debug your MCP server</li></ul><h3>🏥 Understanding MCP with a Hospital Example</h3><p>Imagine you’re visiting a hospital for the first time.</p><p>There are many departments and doctors:</p><ul><li>👨‍⚕️ Cardiologist</li><li>🦴 Orthopedic</li><li>👁️ Eye Specialist</li><li>🩺 General Physician</li></ul><h3>❌ Without a Receptionist</h3><p>You enter the hospital and have no idea where to go.</p><p>You walk from room to room asking people:</p><ul><li>“Is the cardiologist here?”</li><li>“Where is the eye doctor?”</li><li>“Is this the orthopedic department?”</li></ul><p>This takes time, and you may end up in the wrong place.</p><h3>✅ With a Receptionist</h3><p>Now imagine there’s a receptionist at the entrance.</p><p>You simply say:</p><blockquote><em>“I have chest pain. Which doctor should I see?”</em></blockquote><p>The receptionist immediately tells you:</p><blockquote><em>“Please visit Room 205. Dr. Smith, the cardiologist, is available.”</em></blockquote><p>You don’t need to know the hospital layout or search for the right doctor yourself — the receptionist finds the correct destination for you.</p><h3>🤖 How This Relates to MCP</h3><p>Think of the AI assistant as <strong>you</strong>, the user visiting the hospital.</p><p>Think of your Python functions (such as add(), sub(), database_search(), or send_email()) as the <strong>doctors</strong> in different rooms.</p><p>The <strong>MCP Server</strong> acts like the <strong>hospital receptionist</strong>.</p><p>Instead of the AI needing to know where every function is or how to call it, the AI simply asks the MCP server:</p><blockquote><em>“Which tool can help me with this task?”</em></blockquote><p>The MCP server locates the correct tool and executes it.</p><pre>User<br>   │<br>   ▼<br>AI Assistant<br>   │<br>   ▼<br>🏥 MCP Server (Receptionist)<br>   │<br>   ├── ➕ add()<br>   ├── ➖ sub()<br>   ├── ✖️ mul()<br>   └── ➗ divide()</pre><p>The AI doesn’t have to know where each tool lives or how it’s implemented. It simply asks the MCP server, just like a patient asks the receptionist for directions.</p><p><strong>In short:</strong></p><ul><li>👤 <strong>Patient</strong> → User</li><li>🤖 <strong>AI Assistant</strong> → Visitor asking for help</li><li>🏥 <strong>Receptionist</strong> → MCP Server</li><li>👨‍⚕️ <strong>Doctors</strong> → Your Python tools</li><li>🚪 <strong>Hospital Rooms</strong> → Individual functions/services</li></ul><p>This is exactly why MCP is useful — it provides a single, standardized way for AI assistants to discover and use the right tools without needing to know all the implementation details.</p><h3>💡 Tip</h3><blockquote><strong><em>Keep the FastMCP server running</em></strong><em> while using the MCP Inspector. If you stop the server, the Inspector won’t be able to communicate with your tools.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yLA9RiuyF0Kc4HkiyvB7mw.png" /></figure><h3>🎉 You’re Ready!</h3><p>You now have:</p><p>✅ FastMCP Server Running<br>✅ MCP Inspector Connected<br>✅ Interactive Tool Testing Enabled</p><p>You can now invoke tools like <strong>add</strong>, <strong>sub</strong>, <strong>mul</strong>, and <strong>divide</strong> directly from the Inspector UI without writing any client code.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ca3ccf98859c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building a Local AI-Powered Document Intelligence Server with Flask, Ollama & RAG]]></title>
            <link>https://medium.com/@aravinthc18/building-a-local-ai-powered-document-intelligence-server-with-flask-ollama-rag-19b69625b537?source=rss-87e95b073259------2</link>
            <guid isPermaLink="false">https://medium.com/p/19b69625b537</guid>
            <category><![CDATA[chromadb]]></category>
            <category><![CDATA[ollama]]></category>
            <category><![CDATA[easyocr]]></category>
            <category><![CDATA[pytesseract]]></category>
            <category><![CDATA[retrieval-augmented-gen]]></category>
            <dc:creator><![CDATA[aravinth C]]></dc:creator>
            <pubDate>Sun, 14 Dec 2025 06:34:15 GMT</pubDate>
            <atom:updated>2025-12-14T06:37:48.238Z</atom:updated>
            <content:encoded><![CDATA[<h3>From PDFs to Intelligence — all running locally.<em> 🧠💻</em></h3><p>In a world dominated by cloud-based AI services, I wanted to answer a simple question:</p><p><strong>Can we build a complete, production-ready Document AI system that runs fully on local infrastructure?</strong></p><p>This curiosity led to the creation of <strong>LocalDSServer</strong> — a <strong>single Flask-based application</strong> that brings together <strong>LLMs, RAG, OCR, PDF processing, and future ML training pipelines</strong> under one roof.</p><h3>🧩 What is LocalDSServer?</h3><p><strong>LocalDSServer</strong> is a lightweight yet powerful <strong>local document processing and AI inference server</strong>, designed for privacy-first environments like healthcare, legal, and enterprise workflows.</p><p>✨ <strong>No cloud dependency</strong><br>✨ <strong>Runs entirely on local machines</strong><br>✨ <strong>Modular, extensible, and future-ready</strong></p><h3>⚙️ Architecture at a Glance</h3><pre>📄 PDF / Image<br>     ↓<br>🖼 Image Preprocessing<br>     ↓<br>🔍 OCR Engine (pytessearct/easyocr)<br>     ↓<br>📚 RAG + Vector DB (ChromaDB)<br>     ↓<br>🤖 Local LLM (Ollama)<br>     ↓<br>📊 Structured Output / Re-edited PDF</pre><p>All of this is orchestrated through a <strong>single Flask application</strong> — simple, debuggable, and scalable.</p><h3>🤖 Local LLM Integration with Ollama</h3><p>Instead of relying on external APIs, this project integrates <strong>Ollama</strong> to run <strong>LLMs locally</strong>.</p><p>✔️ Fast inference<br>✔️ No data leakage<br>✔️ Full control over prompts and outputs</p><p>This makes the system ideal for <strong>sensitive documents</strong> such as medical records, insurance claims, and internal enterprise files.</p><h3>📚 RAG (Retrieval-Augmented Generation) — Done Right</h3><p>The heart of the system is its <strong>RAG pipeline</strong>.</p><p>🔹 Documents are embedded and stored using <strong>ChromaDB</strong><br>🔹 Embeddings are saved locally for <strong>fast retrieval</strong><br>🔹 Relevant chunks are fetched before every LLM call</p><blockquote><em>⚡ Result: Accurate, context-aware answers without hallucination.</em></blockquote><p>Even as the dataset grows, retrieval stays efficient — keeping the system <strong>on track and performant</strong>.</p><h3>🔍 OCR That Actually Works (Even for Tough PDFs)</h3><p>The OCR pipeline is <strong>multi-engine and adaptive</strong>:</p><p>🧠 <strong>Tesseract (pytesseract)</strong><br>🧠 <strong>EasyOCR</strong></p><p>Before OCR, images go through <strong>preprocessing techniques</strong> like:</p><ul><li>🖤 Grayscale conversion</li><li>🌫 Gaussian blur</li><li>🔆 Contrast enhancement</li><li>📐 Noise reduction</li></ul><p>This dramatically improves text extraction accuracy from:</p><ul><li>Scanned PDFs</li><li>Low-quality images</li><li>Medical reports &amp; forms</li></ul><h3>📝 Re-Editing &amp; Rebuilding PDFs</h3><p>Once text is extracted and structured:</p><p>📌 Content can be <strong>cleaned, reorganized, and re-edited</strong><br>📌 New PDFs can be generated programmatically<br>📌 Ideal for <strong>medical summaries, reports, and compliance documents</strong></p><h3>🧠 What’s Coming Next? (Future Roadmap)</h3><p>This project is not just about inference — it’s about <strong>learning systems</strong> and <strong>agent-driven workflows</strong>.</p><p>🚧 <strong>Upcoming Enhancements</strong>:</p><ul><li>🔗 <strong>LangChain Integration</strong></li><li>Orchestrate multi-step LLM workflows</li><li>Build document-aware chains and agents</li><li>Enable tool calling for OCR, RAG, and PDF actions</li><li>Create reusable pipelines for summarization, extraction, and validation</li><li>🧬 <strong>NER Training Pipeline</strong></li><li>Train Named Entity Recognition models using extracted datasets</li><li>📈 <strong>SVM Models</strong></li><li>Classical ML for classification &amp; document tagging</li><li>🏷 Auto-labeling from RAG + LangChain outputs</li><li>🧪 Dataset-driven fine-tuning workflows</li></ul><p>The long-term goal is to turn <strong>unstructured documents into structured intelligence</strong>.</p><h3>🛠 Why This Project Matters</h3><p>✅ Privacy-first AI<br>✅ Fully offline capability<br>✅ Real-world document workflows<br>✅ Designed for scalability &amp; ML evolution</p><p>This is not a demo — it’s a <strong>foundation</strong> for enterprise-grade Document AI systems.</p><h3>🔗 Open Source &amp; Ready to Explore</h3><p>If you’re interested in:</p><ul><li>Local LLMs</li><li>RAG systems</li><li>LangChain workflows</li><li>OCR pipelines</li><li>Document Intelligence</li></ul><p>👉 <strong>Check out the repository and explore the codebase</strong></p><p>🐙 <strong>Github Repo:</strong><a href="https://github.com/postboxat18/LocalDSServer">LocalDSServer</a></p><p>👉 Fork it, extend it, or build your own AI server on top of it</p><h3>🌟 Final Thoughts</h3><blockquote>AI doesn’t always need the cloud.<em><br></em>Sometimes, the smartest systems live right where your data is.</blockquote><p>Building <strong>LocalDSServer</strong> has been a journey through <strong>LLMs, OCR, RAG, LangChain, and classical ML</strong> — and it’s just getting started.</p><p>If you’re passionate about <strong>local AI, privacy, and real-world ML systems</strong>, this project is for you. 💡🔥</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=19b69625b537" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[✨ Group Similar Items in Python Using Fuzzy Matching (rapidfuzz)]]></title>
            <link>https://medium.com/@aravinthc18/group-similar-items-in-python-using-fuzzy-matching-rapidfuzz-8de358d841af?source=rss-87e95b073259------2</link>
            <guid isPermaLink="false">https://medium.com/p/8de358d841af</guid>
            <category><![CDATA[fuzzy-matching]]></category>
            <category><![CDATA[python-programming]]></category>
            <category><![CDATA[group-by]]></category>
            <dc:creator><![CDATA[aravinth C]]></dc:creator>
            <pubDate>Sun, 13 Jul 2025 08:09:56 GMT</pubDate>
            <atom:updated>2025-10-25T05:52:30.582Z</atom:updated>
            <content:encoded><![CDATA[<p>📌 <strong>Author:</strong> postboxat18<br> 🐙 <strong>GitHub Repo:</strong> <a href="https://github.com/postboxat18/groupbyScore.git">postboxat18/groupbyScore</a></p><blockquote><em>🗂️ Have you ever worked with messy data where strings like </em><em>&quot;Chest Pain&quot; and </em><em>&quot;chest pain&quot; or </em><em>&quot;fever&quot; and </em><em>&quot;feverish&quot; needed to be grouped together?</em></blockquote><p>You’re not alone. Real-world datasets are often noisy — misspellings, inconsistent casing, and slight variations are everywhere. Today, I’ll show you a clean and powerful way to <strong>group similar records</strong> using <strong>fuzzy matching</strong> in Python.</p><p>Let’s dive in. 🏊‍♂️</p><h3>🔍 What Is Fuzzy Grouping?</h3><p>Fuzzy grouping means <strong>clustering similar items</strong> based on their “closeness” or <strong>similarity score</strong>, rather than exact match.</p><p>Instead of writing complex matching logic, we can use the 🔥 rapidfuzz library — a faster alternative to fuzzywuzzy.</p><h3>🚀 The Function: groupbyScore</h3><p>Here’s the custom function we’ll use to group items:</p><pre>from rapidfuzz import fuzz<br>def groupbyScore(all_data,score ,key):<br>    page_list = []<br>    used_list = []<br>    for i, data in enumerate(all_data):<br>        if not i in used_list:<br>            lis = []<br>            for j, sec_data in enumerate(all_data):<br>                isTrue = []<br>                for k in range(len(key(all_data[0]))):<br>                    first_text = key(data)[k]<br>                    sec_text = key(sec_data)[k]<br>                    if isinstance(first_text, str) and isinstance(sec_text, str):<br>                        isTrue.append(fuzz.partial_ratio(first_text.lower(), sec_text.lower()) &gt; 80)<br>                    elif isinstance(first_text, int) and isinstance(sec_text, int):<br>                        isTrue.append(first_text == sec_text)<br>                if all(isTrue) and i not in page_list and j not in page_list:<br>                    lis.append(j)<br>                    used_list.append(j)<br>            non_page_list = sorted(set(lis))<br>            if len(non_page_list) &gt; 1:<br>                page_list.extend(non_page_list)<br>                all_page_list = [all_data[pg] for pg in non_page_list]<br>                yield key(data), all_page_list<br>            else:<br>                yield key(data), [data]</pre><h3>🧪 Try It in Action</h3><p>Let’s use the function on some real-world-like messy data:</p><pre>data = [<br>    {&quot;Notes&quot;: &quot;Chest Pain&quot;, &quot;page_num&quot;: 1},<br>    {&quot;Notes&quot;: &quot;chest pain&quot;, &quot;page_num&quot;: 1},<br>    {&quot;Notes&quot;: &quot;Fever&quot;, &quot;page_num&quot;: 2},<br>    {&quot;Notes&quot;: &quot;feverish&quot;, &quot;page_num&quot;: 2}<br>]<br>for key, group in groupbyScore(data, key=lambda x: (x[&quot;Notes&quot;], x[&quot;page_num&quot;]),score=80):<br>    print(&quot;🔑 Group Key:&quot;, key)<br>    print(&quot;📄 Group Items:&quot;, group)</pre><h3>🖨️ Output:</h3><pre>🔑 Group Key: (&#39;Chest Pain&#39;, 1)<br>📄 Group Items: [{&#39;Notes&#39;: &#39;Chest Pain&#39;, &#39;page_num&#39;: 1}, {&#39;Notes&#39;: &#39;chest pain&#39;, &#39;page_num&#39;: 1}]<br><br>🔑 Group Key: (&#39;Fever&#39;, 2)<br>📄 Group Items: [{&#39;Notes&#39;: &#39;Fever&#39;, &#39;page_num&#39;: 2}, {&#39;Notes&#39;: &#39;feverish&#39;, &#39;page_num&#39;: 2}]</pre><p>✨ Just like that, you’ve clustered similar notes together — even though they weren’t a perfect match.</p><h3>🧰 Why rapidfuzz?</h3><ul><li>🚀 Blazing fast (C++ backend)</li><li>🧠 Smart string similarity functions like partial_ratio, token_set_ratio, etc.</li><li>🛠️ Drop-in replacement for fuzzywuzzy</li></ul><p>Install it using:</p><pre>pip install rapidfuzz</pre><h3>📁 Use Cases</h3><ul><li>Grouping patient notes in healthcare data 🏥</li><li>Deduplicating product listings in e-commerce 🛍️</li><li>Merging survey responses 📋</li><li>Cleaning up user-generated content 🧹</li></ul><h3>💡 Final Thoughts</h3><p>This pattern can save hours of tedious manual cleanup. With just a few lines of code, groupbyScore + rapidfuzz lets you group, clean, and organize your data intelligently.</p><p>Let me know how you’re using it or feel free to contribute to the <a href="https://github.com/postboxat18/groupbyScore.git">GitHub repo</a>! 🧑‍💻</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8de358d841af" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>