<?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 Alper Yıldırım on Medium]]></title>
        <description><![CDATA[Stories by Alper Yıldırım on Medium]]></description>
        <link>https://medium.com/@repfly?source=rss-b488841fe04c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*6U7AFwOm3pdAOpPg6hNCyw.gif</url>
            <title>Stories by Alper Yıldırım on Medium</title>
            <link>https://medium.com/@repfly?source=rss-b488841fe04c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 06:35:38 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@repfly/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[Context Engine for LLMs to Actually Understands Your Codebase]]></title>
            <link>https://repfly.medium.com/context-engine-for-llms-to-actually-understands-your-codebase-90221584730b?source=rss-b488841fe04c------2</link>
            <guid isPermaLink="false">https://medium.com/p/90221584730b</guid>
            <category><![CDATA[retrieval-augmented-gen]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[code]]></category>
            <category><![CDATA[github]]></category>
            <dc:creator><![CDATA[Alper Yıldırım]]></dc:creator>
            <pubDate>Thu, 12 Mar 2026 06:53:06 GMT</pubDate>
            <atom:updated>2026-03-12T06:53:06.880Z</atom:updated>
            <content:encoded><![CDATA[<h4>How Mimir uses code graphs and beam search to give AI the right context, not all the context.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TnTNVMZ8BfaEkKOSnh1qnw.png" /></figure><p>If you’ve ever tried to use Claude or GPT on a large codebase, you already know the pain. You paste in some files, the model hallucinates because it’s missing a dependency three folders away. You paste in more files, now you’re burning tokens on irrelevant code. You try to be smart about what to include, but that manual selection process doesn’t scale. Even in tools like Cursor or Claude Code, because every new chat session is independent, they will try to grep everything with keywords without relevance.</p><p>This is the problem I set out to solve with <strong>Mimir</strong>.</p><h3>The core idea</h3><p>Instead of treating your codebase as a bag of files, Mimir treats it as a graph. Every function, class, and module becomes a node. Every call, type reference, and inheritance relationship becomes an edge. When you ask a question, Mimir doesn’t just search for keywords. It finds the most relevant entry points and then walks the graph to pull in connected code, the stuff that actually matters for answering your question.</p><p>The result is a tight, connected bundle of code that fits within your token budget. No more guessing what to include.</p><h3>How it works under the hood</h3><p>The pipeline has two phases: indexing and retrieval.</p><p>Indexing parses your code with tree-sitter, extracts symbols, and resolves cross-file dependencies automatically. If function A in auth.py calls function B in utils.py, Mimir knows about that relationship. It then generates embeddings for each symbol using local models, so everything runs offline with no API keys needed.</p><p>Retrieval is where it gets interesting. When you ask “how does authentication work?”, Mimir first classifies your intent. Is this a locate query, a trace query, or a debug query? Each type gets different retrieval parameters. Then it runs a hybrid search combining semantic similarity, keyword matching, and name matching. The top results become seeds for a beam search that expands along the graph edges, pulling in callers, callees, type definitions, and config references.</p><p>After expansion, nodes get scored based on retrieval relevance, recency (recently changed code scores higher), change frequency, and connectivity quality. Poorly connected or incomplete nodes get pushed down. Finally, everything is fitted to your token budget by dropping the lowest scoring nodes.</p><h3>What makes it different</h3><p>There are other code search tools out there. Most of them stop at “find relevant files.” Mimir goes further in a few ways:</p><p><strong>Graph-based expansion</strong>. Finding the right function is only half the battle. You usually need its callers, the types it uses, and the config it reads. Mimir follows those edges automatically.</p><p><strong>Session memory.</strong> If you’re having a multi-turn conversation, Mimir tracks what code it already sent to the LLM. It uses an exponential decay model to decide whether to re-send, summarize, or skip previously seen code. This prevents the same 200 lines from appearing in every response.</p><p><strong>Write-path context. </strong>Before you edit a file, you can ask Mimir for write context. It pulls up interfaces the code implements, sibling implementations, test files, and dependency injection registrations. Everything you’d normally hunt for manually.</p><p><strong>Impact analysis.</strong> Before refactoring, you can ask “what breaks if I change this?” Mimir traces callers, type users, and transitive dependencies to show you the blast radius.</p><h3>Multi-repo and team support</h3><p>Mimir handles multiple repositories in a single graph with cross-repo edge detection. For teams, it can run as a shared HTTP server. A central machine indexes all repos, and developers connect through a lightweight client that weighs almost nothing. Mobile developers can query backend code they’ve never cloned. Frontend developers can understand API contracts without switching repos.</p><p>The whole thing runs in Docker too. The embedding model is baked into the image, so it works fully offline.</p><h3>Getting started</h3><pre>pip install mimir-context-server<br>cd /your/project<br>mimir init<br>mimir index<br>mimir serve</pre><p>That’s it. Add the MCP server to your IDE config and your AI assistant now has structured, relevant, graph-aware context for every question.</p><h3>Why I built this</h3><p>Working with LLMs on real codebases shouldn’t require manually curating context windows. The model should get exactly what it needs to answer your question, connected code that respects the actual dependency structure of your project. That’s what Mimir does.</p><p>It’s licensed under MIT, feel free to contribute or give feedback.</p><p>GitHub: <a href="https://github.com/repfly/mimir-context-server">https://github.com/repfly/mimir-context-server</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=90221584730b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What is a Markov Chain]]></title>
            <link>https://repfly.medium.com/what-is-a-markov-chain-5f9cfd21784b?source=rss-b488841fe04c------2</link>
            <guid isPermaLink="false">https://medium.com/p/5f9cfd21784b</guid>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Alper Yıldırım]]></dc:creator>
            <pubDate>Tue, 06 Dec 2022 20:59:44 GMT</pubDate>
            <atom:updated>2022-12-17T19:58:14.472Z</atom:updated>
            <content:encoded><![CDATA[<p>A mathematical system called a Markov chain transitions between states based on a set of probabilistic rules. The distinguishing feature of a Markov chain is that, regardless of how the system got to its current state, the potential future states are fixed. In other words, only the current state and the amount of time have any influence on the likelihood of transitioning to any specific state.</p><p>There are numerous uses for Markov chains, including in natural language processing. A Markov chain can be used to simulate the likelihood of a word sequence in a text in natural language processing. Each word is treated as a state in the chain, and transitions between states are determined by the likelihoods that particular words will appear after other words. As a result, the Markov chain can forecast the likelihood that specific words will appear given the input’s context.</p><p>Take the phrase “The cat sat on the mat,” for instance. We can create the following transition diagram if we think of each word as a state in a Markov chain:</p><pre>  *  --The--&gt;  cat  --sat--&gt;  on  --the--&gt;  mat<br>  |                 |               |<br>  +-----------------+---------------+</pre><p>The arrows in this diagram show the transitions between states, and the weights on the arrows stand in for the probabilities of those transitions. Because the word “cat” always comes after the word “The” in the sentence, the probability of changing from the state “The” to the state “cat” in this instance is 1. Because the word “sat” always comes after the word “cat” in the sentence, the probability of changing from the state “cat” to the state “sat” is also 1.</p><p>This model allows us to determine the probability that any string of words will appear in the sentence. For instance, since the transition from “The” to “cat” has a probability of 1, the probability of the sequence “The cat” occurring is 1. The sequence “The cat sat on” also has a probability of 1 because it is equally likely for “The” to become “cat” and for “cat” to become “sat.”</p><p>Markov chains can be extended to model longer sequences of words, and they can also be trained on large bodies of text data to improve their accuracy. This makes them a powerful tool for natural language processing tasks, such as language modeling and text generation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5f9cfd21784b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Key Points for Performance Optimisation In iOS Apps]]></title>
            <link>https://repfly.medium.com/key-points-for-performance-optimisation-in-ios-apps-8940ace0e172?source=rss-b488841fe04c------2</link>
            <guid isPermaLink="false">https://medium.com/p/8940ace0e172</guid>
            <category><![CDATA[mobile-app-development]]></category>
            <category><![CDATA[swift]]></category>
            <category><![CDATA[performance]]></category>
            <category><![CDATA[ios]]></category>
            <dc:creator><![CDATA[Alper Yıldırım]]></dc:creator>
            <pubDate>Sat, 03 Dec 2022 12:34:30 GMT</pubDate>
            <atom:updated>2022-12-03T12:34:30.687Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*no8mD8sFd7-ck-wpx8s1Ig.jpeg" /></figure><p>Performance optimization is an important aspect of developing iOS applications. A well-optimized app not only provides a smooth user experience, but it can also improve the battery life of the device and reduce data usage.</p><p>There are several techniques and best practices that developers can use to optimize the performance of their iOS app. These include:</p><p>Using the latest version of iOS and the latest version of Xcode. Both of these provide performance improvements and new tools for optimizing app performance. Depending your needs of course.</p><ol><li>Using the Core Animation framework to offload graphics rendering to the device’s GPU. This can improve the performance of graphics-intensive tasks and provide a smoother user experience.</li><li>Using Grand Central Dispatch (GCD) to perform tasks in the background and improve the responsiveness of the app.</li><li>Using Instruments, a tool built into Xcode, to measure and analyze the performance of the app. This can help identify areas where the app is using too much CPU, memory, or other resources.</li><li>Minimizing the use of intensive tasks, such as image decoding and resizing, in the main thread of the app. These tasks can cause the app to become unresponsive and result in a poor user experience.</li><li>Using caching and other optimization techniques to reduce the amount of data that needs to be loaded from the network. This can improve the app’s start-up time and reduce data usage.</li></ol><p>Overall, performance optimization is a crucial aspect of developing iOS applications. By following best practices and using the tools provided by Apple, developers can create apps that provide a smooth and seamless user experience.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8940ace0e172" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Binary search tree algorithm implemented in TypeScript]]></title>
            <link>https://repfly.medium.com/binary-search-tree-algorithm-implemented-in-typescript-eab32b48dc59?source=rss-b488841fe04c------2</link>
            <guid isPermaLink="false">https://medium.com/p/eab32b48dc59</guid>
            <category><![CDATA[typescript]]></category>
            <category><![CDATA[binary-search-tree]]></category>
            <dc:creator><![CDATA[Alper Yıldırım]]></dc:creator>
            <pubDate>Fri, 02 Dec 2022 20:41:11 GMT</pubDate>
            <atom:updated>2022-12-04T17:24:52.417Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l2HxxiKs-iH226fL0YFovg.png" /></figure><p>This class defines a BinarySearchTree with a value property, a left property representing the left subtree, and a right property representing the right subtree. It also includes methods for inserting, searching, and removing elements from the tree, as well as methods for finding the minimum and maximum elements in the tree or a subtree.</p><pre>class BinarySearchTree&lt;T&gt; {<br>  value: T;<br>  left: BinarySearchTree&lt;T&gt; | null;<br>  right: BinarySearchTree&lt;T&gt; | null;<br><br>  constructor(value: T) {<br>    this.value = value;<br>    this.left = null;<br>    this.right = null;<br>  }<br><br>  // Insert a new element into the tree<br>  insert(value: T) {<br>    if (value &lt; this.value) {<br>      if (this.left) {<br>        this.left.insert(value);<br>      } else {<br>        this.left = new BinarySearchTree(value);<br>      }<br>    } else {<br>      if (this.right) {<br>        this.right.insert(value);<br>      } else {<br>        this.right = new BinarySearchTree(value);<br>      }<br>    }<br>  }<br><br>  // Search for a value in the tree<br>  search(value: T): BinarySearchTree&lt;T&gt; | null {<br>    if (value &lt; this.value) {<br>      if (this.left) {<br>        return this.left.search(value);<br>      } else {<br>        return null;<br>      }<br>    } else if (value &gt; this.value) {<br>      if (this.right) {<br>        return this.right.search(value);<br>      } else {<br>        return null;<br>      }<br>    } else {<br>      return this;<br>    }<br>  }<br><br>  // Remove an element from the tree<br>  remove(): BinarySearchTree&lt;T&gt; | null {<br>    // Check if the node is a leaf<br>    if (!this.left &amp;&amp; !this.right) {<br>      return null;<br>    } else if (this.left &amp;&amp; this.right) {<br>      // The node has two children, find the next largest value<br>      const successor = this.right.minimum();<br>      this.value = successor.value;<br>      successor.remove();<br>    } else {<br>      // Check if the node has only one child<br>      if (this.left) {<br>        return this.left;<br>      } else {<br>        return this.right;<br>      }<br>    }<br><br>    return this;<br>  }<br><br>  // Find the node with the minimum value in the subtree<br>  minimum(): BinarySearchTree&lt;T&gt; {<br>    let node = this;<br>    while (node.left) {<br>      node = node.left;<br>    }<br>    return node;<br>  }<br><br>  // Find the node with the maximum value in the subtree<br>  maximum(): BinarySearchTree&lt;T&gt; {<br>    let node = this;<br>    while (node.right) {<br>      node = node.right;<br>    }<br>    return node;<br>  }<br>}</pre><p>You can use this class by creating a new BinarySearchTree instance and calling its methods, like this:</p><pre>const tree = new BinarySearchTree(10);<br>tree.insert(5);</pre><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eab32b48dc59" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to create a QR Code and share it in Flutter]]></title>
            <link>https://repfly.medium.com/how-to-create-a-qr-code-and-share-it-in-flutter-67cc3b89fe61?source=rss-b488841fe04c------2</link>
            <guid isPermaLink="false">https://medium.com/p/67cc3b89fe61</guid>
            <category><![CDATA[qr-code]]></category>
            <category><![CDATA[flutter]]></category>
            <category><![CDATA[flutter-app-development]]></category>
            <dc:creator><![CDATA[Alper Yıldırım]]></dc:creator>
            <pubDate>Sun, 11 Sep 2022 14:58:31 GMT</pubDate>
            <atom:updated>2025-06-22T16:04:21.886Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/450/1*7opHPBHNVCKwqc4VpXpxdA.png" /></figure><h3>Introduction</h3><p>Hello everyone! I hope everybody having a great day. After this article, you will be able to generate QR codes using Flutter and share them in Android, iOS, MacOS and Windows devices. If you want to go straight to the point, full code is at the bottom.</p><h4>Dependencies</h4><p>We will be using these dependencies to make our lives easier.</p><pre><strong>share_plus: ^4.1.0</strong></pre><pre><strong>path_provider: ^2.0.11</strong></pre><pre><strong>qr_flutter: ^4.0.0</strong></pre><p><a href="https://pub.dev/packages/share_plus">share_plus</a> is a Flutter community package that allows you to easily implement native share functionality in your Flutter application.</p><p><a href="https://pub.dev/packages/path_provider">path_provider</a> is a package that developed by Flutter devs. It helps you to access your platforms directories.</p><p><a href="https://pub.dev/packages/qr_flutter">qr_flutter</a> is developed by an independent developer and it will be our star of the show.</p><h3>Generating the QR code</h3><p>First, you will be needing a string to generate code from. After you put your data to the <strong>data</strong> parameter in string format, validate function will generate the QR code and return it with error checking results.</p><pre>final qrValidationResult = QrValidator.validate(</pre><pre>  data: dataToValidateInString,</pre><pre>  version: QrVersions.auto,</pre><pre>  errorCorrectionLevel: QrErrorCorrectLevel.L,</pre><pre>);</pre><p>Then after null checks are done according to your taste, by using <strong>QrPainter</strong>, we’ll create the <strong>QRPainter</strong> object to create an image later.</p><pre>final QRPainter qrPainter = QrPainter.withQr(</pre><pre>  qr: qrCode, // QR object to create</pre><pre>  color: Colors.white, //defines the color of squares</pre><pre>  // embeddedImage: image, you can also embed an image to center of the code</pre><pre>);</pre><p>flutter_qr’s API also allows us to directly generate an image from <strong>QRPainter object</strong> but note that it returns an <strong>Image from dart:ui library</strong>, <strong>not</strong> the <strong>widget</strong>. We will be utilising <strong>toImageData </strong>function since it returns ByteData and allows us to use it easier.</p><pre>final ByteData? bytes = await qrPainter.toImageData(</pre><pre>  200,</pre><pre>  format: ImageByteFormat.png,</pre><pre>);</pre><p>In order to use the share function, platforms require a local path from the device. Therefore, we must write our bytes to the device’s temporary directory. <a href="https://pub.dev/packages/path_provider">Path provider</a> arrives to assist us at this point. We will access the temporary directory in type of Directory by using the package’s getTemporaryDirectory API. The directory of the image will be defined using File class from dart:io, and it will then be generated by invoking the create function. Writing a file requires the file to be in Uint8List format. So we parse into the required format and write the file data to created file.</p><pre>final Directory tempDir = await getTemporaryDirectory();<br>final File tempQrFile = await File(&#39;${tempDir.path}/qr.jpg&#39;).create();</pre><pre>final Uint8List list = bytes.buffer.asUint8List();<br>tempQrFile.writeAsBytesSync(list);</pre><p>In here I wanted users to share the QR code as JPG file. To achieve that simply write the extension and filename in the path.</p><h3>Sharing it</h3><p>Finally, share_plus package comes to help by calling native share functions with given parameters. I did also put the mime type of my file just in case.</p><pre>await Share.shareFiles([tempQrFile.path], mimeTypes: [&#39;image/jpeg&#39;]);</pre><h3>DONE 🎉 🎉 🎉</h3><figure><img alt="sponge bob dusting his hands with text in the top writing “done and done”" src="https://cdn-images-1.medium.com/max/498/1*Zio7Z6Q2JYEHyEoFaum7dA.gif" /></figure><h4>You can check the full code below</h4><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/09a2e5dc5270a824757c01bdf9c94eed/href">https://medium.com/media/09a2e5dc5270a824757c01bdf9c94eed/href</a></iframe><p>Have a great day!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=67cc3b89fe61" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>