<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>Posts on Ahmed Magdy</title>
    <link>https://magdi.work/posts/</link>
    <description>Recent content in Posts on Ahmed Magdy</description>
    <generator>Hugo</generator>
    <language>en-us</language>
    <lastBuildDate>Sun, 30 Aug 2026 00:00:00 +0000</lastBuildDate>
    <atom:link href="https://magdi.work/posts/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Building photo-dedupe: a local-first burst-photo cleaner in Go</title>
      <link>https://magdi.work/posts/photo-dedupe/</link>
      <pubDate>Sun, 30 Aug 2026 00:00:00 +0000</pubDate>
      <guid>https://magdi.work/posts/photo-dedupe/</guid>
      <description>&lt;p&gt;Burst mode leaves you with seven nearly-identical frames and no easy way to pick &amp;ldquo;the good one.&amp;rdquo; &lt;code&gt;photo-dedupe&lt;/code&gt; is a small offline Go CLI (plus an optional local web UI) that does that triage: cluster by capture time, group the shots that actually look alike, keep the sharpest and highest-resolution frame. Nothing is ever deleted — losers move to a quarantine folder for you to review. What&amp;rsquo;s worth writing up isn&amp;rsquo;t the algorithms, which are well-known, but how cleanly the problem splits into independent, testable stages, and how much of the design is really about safety: a readable JSON plan, a move instead of a delete, a one-command restore.&lt;/p&gt;</description>
    </item>
    <item>
      <title>pdf2epub: Building a Self-Hosted PDF→Kindle Pipeline</title>
      <link>https://magdi.work/posts/pdf2epub/</link>
      <pubDate>Thu, 13 Aug 2026 00:00:00 +0000</pubDate>
      <guid>https://magdi.work/posts/pdf2epub/</guid>
      <description>&lt;p&gt;&lt;em&gt;A Go CLI that turns scanned, image-only PDFs into Kindle-ready reflowable EPUBs — and everything that went sideways along the way.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;I had a real problem to solve: a scanned copy of an old Arabic memoir — 420 pages, image-only PDF, no text layer at all. I wanted to read it on a Kindle, reflowable, in Arabic, with proper chapters. Nothing off-the-shelf did that well without uploading a personal scan to some cloud OCR service, which didn&amp;rsquo;t sit right with me.&lt;/p&gt;</description>
    </item>
    <item>
      <title>Building a Local LLM Inference API for Semantic Similarity</title>
      <link>https://magdi.work/posts/llm-inference/</link>
      <pubDate>Sun, 05 Oct 2025 00:00:00 +0000</pubDate>
      <guid>https://magdi.work/posts/llm-inference/</guid>
      <description>&lt;p&gt;At &lt;strong&gt;&lt;a href=&#34;https://whatisgoing.com&#34;&gt;whatisgoing.com&lt;/a&gt;&lt;/strong&gt;, we needed a reliable way to &lt;strong&gt;extract named entities&lt;/strong&gt; from multilingual news articles — especially Arabic news — where the same entity can appear in different linguistic forms.&lt;/p&gt;
&lt;p&gt;For example:&lt;/p&gt;
&lt;p&gt;&amp;ldquo;The Ukraine War&amp;rdquo; and &amp;ldquo;The Russian Invasion&amp;rdquo; both refer to the same geopolitical event — the ongoing conflict in Eastern Europe — yet share no overlapping words. Traditional keyword or similarity-based methods fail to capture such &lt;strong&gt;semantic equivalence&lt;/strong&gt;.&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
