<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0"><channel>
<title>Constantine Dzik — Blog</title>
<link>https://constantinedzik.com</link>
<description>Constantine Dzik — technology lead in practical AI and machine learning: entity resolution, LLM systems, computer vision for manufacturing, and building an AI implementation company.</description>
<language>en-US</language>
<atom:link xmlns:atom="http://www.w3.org/2005/Atom" href="https://constantinedzik.com/blog/rss.xml" rel="self" type="application/rss+xml"/>

  <item>
    <title>How Kimi K3 Scales: MoE, KDA, and Attention Residuals</title>
    <link>https://constantinedzik.com/blog/how-kimi-k3-scales/</link>
    <guid>https://constantinedzik.com/blog/how-kimi-k3-scales/</guid>
    <pubDate>Sun, 26 Jul 2026 12:00:00 GMT</pubDate>
    <description>Kimi K3 combines a sparse mixture of experts, linear attention, and depth-wise routing. Here is what its architecture changes for AI teams.</description>
  </item>
  <item>
    <title>Finding failing solar panels with a digital twin</title>
    <link>https://constantinedzik.com/blog/finding-failing-solar-panels/</link>
    <guid>https://constantinedzik.com/blog/finding-failing-solar-panels/</guid>
    <pubDate>Mon, 13 Jul 2026 12:00:00 GMT</pubDate>
    <description>A practical method for detecting solar-panel anomalies by calibrating a two-diode digital twin from voltage, current, temperature, and irradiance telemetry.</description>
  </item>
</channel></rss>