<?xml version="1.0" encoding="UTF-8"?>
<!--
  RSS feed for the blog.

  This is just a list of your posts in XML. Readers subscribe to
  https://kethavath.com/blog/feed.xml in an app like Feedly / NetNewsWire /
  Reeder, and their app checks this file for new <item> entries.

  TO ADD A POST: copy one <item> block, put it at the TOP of the list, and fill in:
    <title>        the post title
    <link>/<guid>  the full URL of the post (guid must be unique and never change)
    <pubDate>      RFC-822 date. Format: Wed, 01 Apr 2026 00:00:00 +0000
    <description>  the excerpt (plain text)
  Also bump <lastBuildDate> below to the same date.
-->
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Harshith Kethavath — Blog</title>
    <link>https://kethavath.com/blog/</link>
    <description>Writing by Harshith Kethavath on ML research, infrastructure, and things worth thinking about.</description>
    <language>en-us</language>
    <lastBuildDate>Wed, 01 Apr 2026 00:00:00 +0000</lastBuildDate>
    <atom:link href="https://kethavath.com/blog/feed.xml" rel="self" type="application/rss+xml" />

    <item>
      <title>What Cloud Segmentation Taught Us About Vision-Language Models</title>
      <link>https://kethavath.com/blog/posts/cloudprompts.html</link>
      <guid isPermaLink="true">https://kethavath.com/blog/posts/cloudprompts.html</guid>
      <pubDate>Wed, 01 Apr 2026 00:00:00 +0000</pubDate>
      <category>ML Research</category>
      <description>We tested whether prompt engineering can adapt vision-language models to satellite imagery. Across 60 prompt variants on cloud segmentation, every one underperformed the zero-shot baseline. But fine-tuning with just 8 labeled images beat them all, and the choice between LoRA and full fine-tuning turned out to be about task structure, not compute.</description>
    </item>

  </channel>
</rss>
