<?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 kalpit borkar</title><link>https://kalpitborkar.github.io/posts/</link><description>Recent content in Posts on kalpit borkar</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Sun, 17 May 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://kalpitborkar.github.io/posts/index.xml" rel="self" type="application/rss+xml"/><item><title>why lab models fail in production</title><link>https://kalpitborkar.github.io/posts/ai-detection-in-the-wild/</link><pubDate>Sun, 17 May 2026 00:00:00 +0000</pubDate><guid>https://kalpitborkar.github.io/posts/ai-detection-in-the-wild/</guid><description>&lt;p&gt;nothing humbles a 99% accurate ai detection model like a screenshot sent over whatsapp.&lt;/p&gt;
&lt;p&gt;in the lab, things are a controlled sandbox. we train on crisp clean data where we know the exact generator, the resolutions, and the baseline jpeg compressions.&lt;/p&gt;
&lt;p&gt;in the wild, things are more chaotic. we deal with screenshots of screenshots, multiple social media preprocessing, arbitrary noise, and classic photoshop retouches. we don’t know the source, and we definitely don&amp;rsquo;t know the generator.&lt;/p&gt;</description></item></channel></rss>