<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Zenzer</title>
    <description>The latest articles on DEV Community by Zenzer (@zenzerjs).</description>
    <link>https://dev.to/zenzerjs</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4246544%2F9f87044d-f313-4b88-b480-499dbb16718a.jpg</url>
      <title>DEV Community: Zenzer</title>
      <link>https://dev.to/zenzerjs</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/zenzerjs"/>
    <language>en</language>
    <item>
      <title>Why LLMs Quietly Round Up Your Resume Metrics (And How I Blocked It Without Another LLM)</title>
      <dc:creator>Zenzer</dc:creator>
      <pubDate>Sun, 15 Nov 2026 20:39:31 +0000</pubDate>
      <link>https://dev.to/zenzerjs/why-llms-quietly-round-up-your-resume-metrics-and-how-i-blocked-it-without-another-llm-2fhk</link>
      <guid>https://dev.to/zenzerjs/why-llms-quietly-round-up-your-resume-metrics-and-how-i-blocked-it-without-another-llm-2fhk</guid>
      <description>&lt;p&gt;If you ask an AI to "tailor my resume to this job description without making anything up," it still tries to help you a little too much.&lt;/p&gt;

&lt;p&gt;While building an open-source resume editor, I kept seeing models do things like:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Rounding a &lt;code&gt;38%&lt;/code&gt; speedup up to &lt;code&gt;40%&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;Dropping &lt;code&gt;Intern&lt;/code&gt; from &lt;code&gt;Software Engineer Intern&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;Taking a real number (&lt;code&gt;10M requests&lt;/code&gt;) and attaching it to something else (&lt;code&gt;10M users&lt;/code&gt;)&lt;/li&gt;
&lt;li&gt;Stretching an internship date range by an extra month&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;None of these look like obvious lies when you skim the output. You usually only find out when an interviewer asks a follow-up question about a number you never actually hit.&lt;/p&gt;

&lt;p&gt;After getting ghosted through fall recruiting as a second-year CS student, I built &lt;strong&gt;&lt;a href="https://resumeforge.studio/" rel="noopener noreferrer"&gt;ResumeForge&lt;/a&gt;&lt;/strong&gt; (&lt;a href="https://github.com/ZenzerJs/ResumeForge" rel="noopener noreferrer"&gt;GitHub, MIT&lt;/a&gt;). Here is how I stopped the AI from sneaking in exaggerated edits before anything can apply or export to PDF.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why "don't lie" in the prompt isn't enough
&lt;/h2&gt;

&lt;p&gt;Telling the AI to stick to the truth helps, and in ResumeForge every AI edit has to link back to an &lt;strong&gt;Evidence Bank&lt;/strong&gt; of real bullets you added—or flag an honest &lt;strong&gt;Gap&lt;/strong&gt; when you're missing a skill the job wants.&lt;/p&gt;

&lt;p&gt;Still, prompts alone don't stop models from rounding &lt;code&gt;38%&lt;/code&gt; to &lt;code&gt;40%&lt;/code&gt;. And using a second AI call to check the first one adds wait time, costs extra tokens, and can miss the exact same small edits.&lt;/p&gt;

&lt;p&gt;Instead, every AI suggestion runs through a plain code check against your original resume and Evidence Bank before you can apply or export it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the code check actually looks for
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. Rounded numbers and swapped topics
&lt;/h3&gt;

&lt;p&gt;First, the checker pulls out every number, its unit (&lt;code&gt;%&lt;/code&gt;, &lt;code&gt;ms&lt;/code&gt;, &lt;code&gt;s&lt;/code&gt;, &lt;code&gt;k&lt;/code&gt;, &lt;code&gt;M&lt;/code&gt;, &lt;code&gt;$&lt;/code&gt;, &lt;code&gt;users&lt;/code&gt;, etc.), and the word right after it from both your original resume and the AI's new version.&lt;/p&gt;

&lt;p&gt;Then it checks that the exact number and unit actually exist in your real experience:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;valueUnitMatches&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;masterFacts&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;metrics&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;filter&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;m&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;m&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;isTrivial&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;sameUnit&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;m&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;unit&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;toLowerCase&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="n"&gt;candMetric&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;unit&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;toLowerCase&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;sameValue&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;Math&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;abs&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;m&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;value&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;candMetric&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;value&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="mf"&gt;0.001&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;sameUnit&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="n"&gt;sameValue&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;valueUnitMatches&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;length&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="n"&gt;violations&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;push&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="n"&gt;kind&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"metric"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;severity&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"HARD"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;field&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"metrics"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;message&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sb"&gt;`Unverified or inflated metric claim "${candMetric.raw}" not supported by master evidence.`&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If &lt;code&gt;38%&lt;/code&gt; turns into &lt;code&gt;40%&lt;/code&gt;, &lt;code&gt;valueUnitMatches&lt;/code&gt; comes back empty and the edit is blocked. It also checks the noun after the number, so the AI can't take a real &lt;code&gt;10M requests&lt;/code&gt; stat and turn it into &lt;code&gt;10M users&lt;/code&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Upgraded job titles
&lt;/h3&gt;

&lt;p&gt;Models love nudging your title closer to the job posting. To block that, the checker compares the words in your title using two simple rules:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The AI can't add new title words that weren't in your original title (which blocks adding &lt;code&gt;Senior&lt;/code&gt;, &lt;code&gt;Staff&lt;/code&gt;, &lt;code&gt;Lead&lt;/code&gt;, etc.).&lt;/li&gt;
&lt;li&gt;If your real title includes words like &lt;code&gt;Intern&lt;/code&gt;, &lt;code&gt;Co-op&lt;/code&gt;, or &lt;code&gt;Junior&lt;/code&gt;, the AI isn't allowed to drop them.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;So shortening &lt;code&gt;Full Stack Software Engineer&lt;/code&gt; to &lt;code&gt;Software Engineer&lt;/code&gt; is allowed, while turning &lt;code&gt;Software Engineer Intern&lt;/code&gt; into &lt;code&gt;Software Engineer&lt;/code&gt; gets blocked.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Stretched dates
&lt;/h3&gt;

&lt;p&gt;Start and end dates (&lt;code&gt;May 2022 - Aug 2023&lt;/code&gt;, &lt;code&gt;Jun 2024 – Present&lt;/code&gt;) are compared against your original resume. The AI can never push a start date earlier, push an end date later, or change a finished job to &lt;code&gt;Present&lt;/code&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Honest gap flags
&lt;/h3&gt;

&lt;p&gt;When a job description asks for a tool you don't have in your resume or Evidence Bank, the AI is required to list it as an &lt;strong&gt;Explicit Gap&lt;/strong&gt; instead of sneaking it into a bullet. If a new skill still shows up in an edit, the app flags it and asks you to confirm whether you actually know it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What still slips through right now
&lt;/h2&gt;

&lt;p&gt;Plain code rules are fast and free, but they still have a few blind spots I'm working on next:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Small plain numbers from 0 to 10 without units&lt;/strong&gt; (like changing &lt;code&gt;team of 3&lt;/code&gt; to &lt;code&gt;team of 5&lt;/code&gt;—skipped right now so normal wording doesn't get falsely blocked, though &lt;code&gt;3x&lt;/code&gt;, &lt;code&gt;5%&lt;/code&gt;, and &lt;code&gt;12&lt;/code&gt; &lt;em&gt;are&lt;/em&gt; checked).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Spelled-out numbers&lt;/strong&gt; like &lt;code&gt;"forty percent"&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Adding a &lt;code&gt;+&lt;/code&gt; sign&lt;/strong&gt; (&lt;code&gt;40%&lt;/code&gt; -&amp;gt; &lt;code&gt;40%+&lt;/code&gt;), since &lt;code&gt;+&lt;/code&gt; currently gets stripped when reading the unit.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Try it out (no login or API key needed)
&lt;/h2&gt;

&lt;p&gt;ResumeForge is built with Next.js, TypeScript, PostgreSQL/Prisma, CodeMirror, and in-browser Typst WASM for live PDF previews. Outside of the pre-loaded demo mode, you can plug in your own API key (BYOK) without your key ever being saved on the server:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Live demo (no signup or API key):&lt;/strong&gt; &lt;a href="https://resumeforge.studio/" rel="noopener noreferrer"&gt;https://resumeforge.studio/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Source code (MIT):&lt;/strong&gt; &lt;a href="https://github.com/ZenzerJs/ResumeForge" rel="noopener noreferrer"&gt;https://github.com/ZenzerJs/ResumeForge&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you spot a clever way an AI could sneak a fake claim past these checks, let me know, I'd love to test it and patch the gap.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>typescript</category>
      <category>nextjs</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
