<?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: Sanketh Kumar</title>
    <description>The latest articles on DEV Community by Sanketh Kumar (@sank8).</description>
    <link>https://dev.to/sank8</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%2F2192163%2Ff047a549-1f5a-4c9d-ad32-3c3c1f0eed8e.jpeg</url>
      <title>DEV Community: Sanketh Kumar</title>
      <link>https://dev.to/sank8</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/sank8"/>
    <language>en</language>
    <item>
      <title>The model wrote six sentences about a grandmother. Two were allowed on the altar.</title>
      <dc:creator>Sanketh Kumar</dc:creator>
      <pubDate>Thu, 29 Oct 2026 03:13:37 +0000</pubDate>
      <link>https://dev.to/sank8/the-model-wrote-six-sentences-about-a-grandmother-two-were-allowed-on-the-altar-18nn</link>
      <guid>https://dev.to/sank8/the-model-wrote-six-sentences-about-a-grandmother-two-were-allowed-on-the-altar-18nn</guid>
      <description>&lt;p&gt;The brief this week listed a "digital ofrenda builder" as one of its ideas. I had been circling the same thought all week, so I took it as permission.&lt;/p&gt;

&lt;p&gt;An &lt;em&gt;ofrenda&lt;/em&gt; is the altar you build for someone who has died: marigolds, candles, a photograph, and the things you want to remember about them. Most of it you arrange by hand. The written part is the part that scares me, because a small language model writing about a real person will invent a detail and hand it to you in exactly the same confident tone as the true ones.&lt;/p&gt;

&lt;p&gt;So I built &lt;strong&gt;ofrenda&lt;/strong&gt;. It writes the remembrance offline with a local open-weight model, and it is not allowed to put a sentence on the page that the memories do not support.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;ollama pull llama3.2:1b
python3 ofrenda.py &lt;span class="nt"&gt;--memories&lt;/span&gt; examples/memories.json &lt;span class="nt"&gt;--out&lt;/span&gt; altar.html
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;One command, one self-contained HTML file, no API key, no account, no network. The six memories in that example file are invented, which is why I can show them here. The tool is for the ones that aren't.&lt;/p&gt;

&lt;h2&gt;
  
  
  The failure I designed against
&lt;/h2&gt;

&lt;p&gt;Ask &lt;code&gt;llama3.2:1b&lt;/code&gt; for a sentence about a grandmother's marigolds and you may get this:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;She kept marigolds in terracotta pots along the kitchen window. &lt;strong&gt;[1]&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Or this:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;She grew up in Oaxaca and kept twelve pots on the sill. &lt;strong&gt;[1]&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The second one is a false memory about someone's grandmother, formatted identically to the first, on a page you were about to send to your family. Hallucination is a nuisance in a chat window. In a memorial it is the entire failure — the one thing the tool cannot do is be confidently wrong about a person who is gone.&lt;/p&gt;

&lt;p&gt;The standard answer is "use a bigger model." I don't think that's an answer. A frontier model invents less often, not never, and it invents with better grammar, which makes it harder to catch. The fix has to be structural.&lt;/p&gt;

&lt;h2&gt;
  
  
  The rule: a sentence must cite the memory it came from
&lt;/h2&gt;

&lt;p&gt;The prompt numbers the memories and asks for a citation on every line — &lt;code&gt;[1]&lt;/code&gt;, &lt;code&gt;[2][3]&lt;/code&gt;. Then Python checks that the citation actually supports what the sentence claims. It looks for the details that turn a general memory into a specific false one:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;unsupported_claims&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;sentence&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;memories&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;person&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="n"&gt;sentence&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;citations&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;no citation&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;                       &lt;span class="c1"&gt;# uncited, gone
&lt;/span&gt;
    &lt;span class="n"&gt;cited&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;by_index&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;i&lt;/span&gt;&lt;span class="p"&gt;].&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;i&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;sentence&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;citations&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;i&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;by_index&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="n"&gt;cited&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;cites memories that do not exist: &lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;sentence&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;citations&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;

    &lt;span class="n"&gt;problems&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[]&lt;/span&gt;
    &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;position&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;word&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="nf"&gt;enumerate&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;sentence&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;split&lt;/span&gt;&lt;span class="p"&gt;()):&lt;/span&gt;
        &lt;span class="n"&gt;cleaned&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;word&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;strip&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;.,;:!?&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;()[]&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;re&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;fullmatch&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;r&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;\d+&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;cleaned&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;             &lt;span class="c1"&gt;# "12"
&lt;/span&gt;            &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="nf"&gt;_supported&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;cleaned&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;cited&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
                &lt;span class="n"&gt;problems&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;number &lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;cleaned&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="k"&gt;continue&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;cleaned&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;lower&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;NUMBER_WORDS&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;           &lt;span class="c1"&gt;# "twelve"
&lt;/span&gt;            &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="nf"&gt;_supported&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;cleaned&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;cited&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="nf"&gt;_supported&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;NUMBER_WORDS&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;cleaned&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;lower&lt;/span&gt;&lt;span class="p"&gt;()],&lt;/span&gt; &lt;span class="n"&gt;cited&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
                &lt;span class="n"&gt;problems&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;number &lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;cleaned&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="k"&gt;continue&lt;/span&gt;

        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;position&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt; &lt;span class="ow"&gt;or&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="n"&gt;cleaned&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;].&lt;/span&gt;&lt;span class="nf"&gt;isupper&lt;/span&gt;&lt;span class="p"&gt;():&lt;/span&gt; &lt;span class="c1"&gt;# not a proper noun
&lt;/span&gt;            &lt;span class="k"&gt;continue&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;cleaned&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;lower&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;SAFE_CAPITALS&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;          &lt;span class="c1"&gt;# "She", "Sunday"...
&lt;/span&gt;            &lt;span class="k"&gt;continue&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;cleaned&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;lower&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;person&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;lower&lt;/span&gt;&lt;span class="p"&gt;():&lt;/span&gt;         &lt;span class="c1"&gt;# her own name
&lt;/span&gt;            &lt;span class="k"&gt;continue&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="nf"&gt;_supported&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;cleaned&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;cited&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
            &lt;span class="n"&gt;problems&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;'"&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;cleaned&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;"'&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;problems&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;_supported&lt;/code&gt; matches case-insensitively as a substring, so a memory that says "marigolds" supports "marigold" and one that says "Sunday" supports "Sundays". Deliberately loose in the direction of not rejecting fair paraphrases, strict about names and numbers.&lt;/p&gt;

&lt;p&gt;Anything that fails gets one repair attempt against only the facts it cited, and then it is dropped.&lt;/p&gt;

&lt;h2&gt;
  
  
  The rule I had to add after watching that one fail
&lt;/h2&gt;

&lt;p&gt;The first version of this passed its own tests and then failed the moment I ran it for real.&lt;/p&gt;

&lt;p&gt;The model wrote six sentences. Two of them were fine. One of them was this:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;She sang while she worked, even when the task seemed impossible. &lt;strong&gt;[4]&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Memory 4 is the one about October candles and a photograph on the table. Memory 3 is the one about singing while she worked. &lt;strong&gt;The sentence was true and the citation was wrong&lt;/strong&gt;, and every check above waved it through: there is no invented name in it, no invented number, and the memory it cited obviously exists. Nothing in the sentence was false. It was just attached to the wrong fact, which for a memorial is the same betrayal wearing a different coat.&lt;/p&gt;

&lt;p&gt;A citation is not a claim of provenance if nothing checks it against the text. So there is a third rule: the sentence has to share at least one content word with the memory it cites.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;claimed&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;content_words&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;sentence&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;claimed&lt;/span&gt; &lt;span class="o"&gt;-=&lt;/span&gt; &lt;span class="nf"&gt;set&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;person&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;lower&lt;/span&gt;&lt;span class="p"&gt;().&lt;/span&gt;&lt;span class="nf"&gt;split&lt;/span&gt;&lt;span class="p"&gt;())&lt;/span&gt;
&lt;span class="n"&gt;supported_words&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;set&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;
&lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;source&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;cited&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;supported_words&lt;/span&gt; &lt;span class="o"&gt;|=&lt;/span&gt; &lt;span class="nf"&gt;content_words&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;source&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;claimed&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="nf"&gt;any&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="nf"&gt;words_match&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;claim&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;word&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;claim&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;claimed&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;word&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;supported_words&lt;/span&gt;
&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;problems&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;NO_SHARED_WORDING&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;content_words&lt;/code&gt; drops stopwords, and two words count as the same when one is a prefix of the other, so "baked" matches "bake" and "singing" matches "sing". Irregular verbs like "sang" do not match — that errs towards refusing, which is the direction I want. Re-running the same prompt, the same model, the same memories, the singing sentence is now refused:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;it cites [4], but shares no wording with that memory&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Two of six sentences survived. The other four are on the page anyway, struck through, with the reason next to each one.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fohfwuozxmnc14hjo5t3e.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fohfwuozxmnc14hjo5t3e.png" alt="The finished altar: two sentences, the six memories they had to come from, and the four the check refused" width="800" height="1167"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The part I care about most: the refusals stay visible
&lt;/h2&gt;

&lt;p&gt;A guard that silently deletes is indistinguishable from a guard that deletes nothing. You cannot audit what you cannot see, and a memorial page is exactly the wrong place to ask for trust.&lt;/p&gt;

&lt;p&gt;So the page has a section called &lt;strong&gt;What it was not allowed to say&lt;/strong&gt;. It shows each refused sentence struck through, and under it the memory it cited and what was wrong. That is where a reader learns that the model produced &lt;em&gt;"Elena's love for life was evident in every detail"&lt;/em&gt; with no citation at all, and that it cited the tin of buttons for a sentence about October.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgthbqeuy0657wrvi7q0n.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgthbqeuy0657wrvi7q0n.png" alt="What it was not allowed to say: four refused sentences and the exact reason for each" width="800" height="599"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Naming the memory in the reason was a late fix, and it matters more than it sounds. My first version said only &lt;code&gt;"October"&lt;/code&gt; was unsupported. The sentence was true — memory 4 &lt;em&gt;does&lt;/em&gt; say October — so the refusal looked like a bug in my tool rather than a mistake by the model. The sentence had cited &lt;code&gt;[5]&lt;/code&gt;, the tin of buttons. Once the reason says &lt;em&gt;which&lt;/em&gt; memory, the page is honest about what happened, and the person reading it can tell the difference between a bad model and a bad guard.&lt;/p&gt;

&lt;h2&gt;
  
  
  And the one it still missed
&lt;/h2&gt;

&lt;p&gt;The second kept sentence is this:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Her kitchen was always filled with the sweet scent of freshly baked goods. &lt;strong&gt;[3]&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Memory 3 says she sang while she worked, in the kitchen, in the garden, even when the work was hard. It does not say anything about baking. The sentence passed because it shares one word — &lt;em&gt;kitchen&lt;/em&gt; — with the memory it cites. The bread is in memory 2, and the model cited the wrong one again.&lt;/p&gt;

&lt;p&gt;I could tighten the rule until this is caught: require most content words to match, or the rarest one. But every step in that direction starts refusing fair paraphrases, and a tool that refuses the true sentence about your grandmother because you phrased it differently is a worse tool. So I stopped, wrote the limit down, and left the miss visible on the page — which is the whole reason the page shows you every memory next to the sentence that used it. &lt;strong&gt;The guard narrows what the model can get away with. Reading the result is still the last line of defence, and the design should make that easy rather than pretend it isn't needed.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Why local, specifically
&lt;/h2&gt;

&lt;p&gt;Three reasons, in the order they mattered to me.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The memories are the most personal text a person owns.&lt;/strong&gt; Sending them to a hosted model means someone else's servers hold a copy of what you remember about your grandmother, under terms nobody reads. That is not a trade I want to make on someone's behalf, and it is not one this tool asks them to make.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It works with the internet off.&lt;/strong&gt; Not as a gimmick — a memorial is something you might want to open in ten years, on a laptop, at a kitchen table, without an account that has since been deleted. The output is one HTML file with no external fonts, scripts, or images. It opens from disk, in ten years, on a machine with nothing installed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;You can inspect the limits of a model you run yourself.&lt;/strong&gt; &lt;code&gt;llama3.2:1b&lt;/code&gt; is 1.3 GB and noticeably worse at prose than a frontier model. That is exactly why this project looks the way it does: the guard exists because I can see how small the model is and I know what it will do. A model behind an API is a model whose failure modes you find out about after you publish.&lt;/p&gt;

&lt;p&gt;Small and checked beats large and trusted, for this job.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'd tell someone building the same thing
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Validate provenance, not format.&lt;/strong&gt; A well-formed citation is a generated token like any other. Mine cited a real memory, in the right syntax, about the wrong thing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The first version of a guard passes its own tests and fails on the first real run.&lt;/strong&gt; Mine did, twice — once for wrong attribution, once for a reason that didn't name the memory and so read as a bug.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Show the refusals.&lt;/strong&gt; It is the difference between a tool you can audit and a tool you have to believe.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Write down what the guard misses.&lt;/strong&gt; Mine misses generic sentiment that shares a word with the memory it cites. Knowing that is more useful than a claim of correctness.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/sank8-2/ofrenda
&lt;span class="nb"&gt;cd &lt;/span&gt;ofrenda
python3 &lt;span class="nt"&gt;-m&lt;/span&gt; unittest &lt;span class="nt"&gt;-v&lt;/span&gt; test_ofrenda     &lt;span class="c"&gt;# 61 tests, no model, no network&lt;/span&gt;

python3 ofrenda.py &lt;span class="nt"&gt;--memories&lt;/span&gt; examples/memories.json &lt;span class="nt"&gt;--out&lt;/span&gt; altar.html &lt;span class="nt"&gt;--dry-run&lt;/span&gt;
python3 ofrenda.py &lt;span class="nt"&gt;--memories&lt;/span&gt; examples/memories.json &lt;span class="nt"&gt;--out&lt;/span&gt; altar.html
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;--dry-run&lt;/code&gt; renders from a canned reply so you can see the layout without waiting, and one sentence in that reply deliberately invents a place and a number so you can watch the verifier refuse a lie.&lt;/p&gt;

&lt;p&gt;On this laptop the first run took 90 seconds and a warm one 13. It is a 1B model on a CPU, and that is the honest cost of keeping the memories at home.&lt;/p&gt;

&lt;p&gt;Code: &lt;strong&gt;&lt;a href="https://github.com/sank8-2/ofrenda" rel="noopener noreferrer"&gt;github.com/sank8-2/ofrenda&lt;/a&gt;&lt;/strong&gt; — MIT, one file, standard library only. Bring your own memories; the tool is more useful that way.&lt;/p&gt;

</description>
      <category>hf26challenge</category>
      <category>opensource</category>
      <category>python</category>
      <category>devchallenge</category>
    </item>
    <item>
      <title>My 1B model renamed one file to another file's name — the fix was a provenance guard</title>
      <dc:creator>Sanketh Kumar</dc:creator>
      <pubDate>Thu, 22 Oct 2026 05:36:41 +0000</pubDate>
      <link>https://dev.to/sank8/my-1b-model-renamed-one-file-to-another-files-name-the-fix-was-a-provenance-guard-3ekm</link>
      <guid>https://dev.to/sank8/my-1b-model-renamed-one-file-to-another-files-name-the-fix-was-a-provenance-guard-3ekm</guid>
      <description>&lt;p&gt;My Downloads folder had &lt;strong&gt;599 loose files in it&lt;/strong&gt; — 101 PDFs, 79 spreadsheets, 129 images, and a folder full of things called &lt;code&gt;img.png&lt;/code&gt;, &lt;code&gt;fp.jpeg&lt;/code&gt; and &lt;code&gt;1762857107887_floop-plan-4.png&lt;/code&gt;. I had been closing that tab for months.&lt;/p&gt;

&lt;p&gt;So for this week's Hacktoberfest challenge ("build something with open-source AI that rewards laziness") I wrote &lt;strong&gt;Tidy Sloth&lt;/strong&gt;: a local file tidier that sorts a mess into folders and gives files names a human can actually read. Python standard library, Ollama, no cloud, no API key.&lt;/p&gt;

&lt;p&gt;It works. But the interesting part isn't the sorting. It's that a 1-billion-parameter model confidently renamed one file to &lt;em&gt;a completely different file's name&lt;/em&gt;, and the fix was not a better prompt.&lt;/p&gt;

&lt;h2&gt;
  
  
  The sloth is the architecture, not the mascot
&lt;/h2&gt;

&lt;p&gt;Sloths are slow because speed costs energy. A file tidier has the same economics: most files are obvious, and asking a language model about them is wasted work.&lt;/p&gt;

&lt;p&gt;So the pipeline spends tokens like they're calories:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;file ──▶ rules ──▶ confident? ──yes──▶ done, 0 tokens
                       │
                       no
                       ▼
              local model (batched) ──▶ category + name
                       │
                       ▼
             Python validates, then moves
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Extension rules&lt;/strong&gt; — &lt;code&gt;png → Images&lt;/code&gt;, &lt;code&gt;xlsx → Data&lt;/code&gt;, &lt;code&gt;zip → Archives&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Purpose rules beat extension&lt;/strong&gt; — &lt;code&gt;Invoice-March.pdf&lt;/code&gt; goes to &lt;code&gt;Invoices&lt;/code&gt;, not &lt;code&gt;Documents&lt;/code&gt;. &lt;code&gt;Resume-Sanketh.pdf&lt;/code&gt; goes to &lt;code&gt;Resumes&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A meaningful filename is enough&lt;/strong&gt; — &lt;code&gt;carprices.csv&lt;/code&gt; doesn't need a model to tell you it's data.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Only the leftovers go to &lt;code&gt;llama3.2:1b&lt;/code&gt;: the timestamps, the &lt;code&gt;img.png&lt;/code&gt;s, the &lt;code&gt;AC3.json&lt;/code&gt;s.&lt;/p&gt;

&lt;p&gt;On my real folder:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;loose files&lt;/td&gt;
&lt;td&gt;599&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;resolved by rules&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;453 (75.6%)&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;needed the model&lt;/td&gt;
&lt;td&gt;146&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;model calls&lt;/td&gt;
&lt;td&gt;13 (batches of 12)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;wall time&lt;/td&gt;
&lt;td&gt;14.2 min&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;estimated manual sorting&lt;/td&gt;
&lt;td&gt;~80 min&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Three quarters of the folder never touched the model. That's the whole trick.&lt;/p&gt;

&lt;h2&gt;
  
  
  Then the model lied
&lt;/h2&gt;

&lt;p&gt;Small models are not editors. I knew that. What I didn't expect was &lt;em&gt;how&lt;/em&gt; it failed.&lt;/p&gt;

&lt;p&gt;I batch a dozen filenames per prompt and ask for a category and a cleaner name for each. The model returned exactly twelve answers, in order, in valid JSON. The count matched. The JSON was clean. And it had silently shifted the associations:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;1786076876258_SoulTree-Brochure-1 (1).pdf  →  AC2.pdf
1786076876258_SoulTree-Brochure-1.pdf      →  Cost-Breakdown-6a61d9864a29092bf3f7159d-1.pdf
1748950125346_out-0.webp                   →  Aadhar-Card.webp
1764919659917_floop-plan-4.png             →  Copy-of-2024-07-TBC-Banner-Generic-2560x650-1-H1nZTAL.png
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Every one of those names belongs to a &lt;em&gt;different file in the same batch&lt;/em&gt;. The model didn't hallucinate a filename; it recycled a real one. Which is worse, because it looks plausible enough to ship.&lt;/p&gt;

&lt;p&gt;My first fix was to stop trusting the model's self-reported index and match answers positionally. That didn't help — the count was already right, so there was nothing to catch. I also stripped echoed extensions, which fixed &lt;code&gt;thing.pdf.pdf&lt;/code&gt; but not this.&lt;/p&gt;

&lt;h2&gt;
  
  
  The fix: a rename may only reuse the file's own words
&lt;/h2&gt;

&lt;p&gt;The rule I settled on is boring and it works. &lt;strong&gt;A proposed name is accepted only if it is built from words that were already in the original filename.&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;provenance_ok&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;proposed&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;original&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="sh"&gt;"""&lt;/span&gt;&lt;span class="s"&gt;A rename may reorganise a file&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;s own words. Never import another&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;s.&lt;/span&gt;&lt;span class="sh"&gt;"""&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="nf"&gt;_tokens&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;proposed&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&lt;/span&gt; &lt;span class="nf"&gt;_tokens&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;original&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;
    &lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;b&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;_compact&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;proposed&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt; &lt;span class="nf"&gt;_compact&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;original&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nf"&gt;bool&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="n"&gt;b&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;b&lt;/span&gt; &lt;span class="ow"&gt;or&lt;/span&gt; &lt;span class="n"&gt;b&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;SoulTree-Brochure-1.pdf&lt;/code&gt; → &lt;code&gt;AC2.pdf&lt;/code&gt; shares zero tokens with the source, so it's rejected and the file keeps its own name. &lt;code&gt;SoulaceNew2.png&lt;/code&gt; → &lt;code&gt;soulace-new-2&lt;/code&gt; shares everything, so it's accepted.&lt;/p&gt;

&lt;p&gt;The model may &lt;strong&gt;reorganise&lt;/strong&gt; a filename. It may never &lt;strong&gt;import&lt;/strong&gt; one.&lt;/p&gt;

&lt;p&gt;I put that and the other invariants into 15 unit tests, including the exact five bad pairs the model produced:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;$ python3 -m unittest test_tidysloth
Ran 15 tests in 0.002s
OK
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;After the guard, the same 599-file run produced &lt;strong&gt;zero cross-row names&lt;/strong&gt; and zero doubled extensions.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two more rules that saved me
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;A name with no meaning is never given invented meaning.&lt;/strong&gt; Given &lt;code&gt;01.csv&lt;/code&gt;, the model happily suggested &lt;code&gt;project-proposal.csv&lt;/code&gt;. That's not tidying, that's fiction — and it's unfalsifiable, so you'd never notice it was wrong. Now &lt;code&gt;01.csv&lt;/code&gt; stays &lt;code&gt;01.csv&lt;/code&gt;. Only the category is decided by the model.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Deterministic noise removal beats model creativity.&lt;/strong&gt; &lt;code&gt;1762857107887_floop-plan-4.png&lt;/code&gt; became &lt;code&gt;floop-plan-4.png&lt;/code&gt; by stripping a leading 10+ digit message ID in Python — no tokens, no guessing, and it works whether or not Ollama is running. (With one edge case: &lt;code&gt;4446563839.pdf&lt;/code&gt; is &lt;em&gt;all&lt;/em&gt; digits, so stripping the "ID" would leave nothing. The strip only applies if a real name survives.)&lt;/p&gt;

&lt;h2&gt;
  
  
  Sorting is destructive, so the defaults are cowards
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;plan&lt;/code&gt; changes &lt;strong&gt;nothing&lt;/strong&gt;. It writes &lt;code&gt;plan.json&lt;/code&gt; and an HTML report you read in the browser first.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;apply&lt;/code&gt; moves files and writes a timestamped manifest.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;undo&lt;/code&gt; replays that manifest in reverse and removes the empty folders it created.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;There is no delete code path.&lt;/strong&gt; Collisions get renamed, never overwritten.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I tested apply and undo on a 40-file copy: 40 files into 10 folders, then all 40 restored and the folders cleaned up. I have deliberately &lt;strong&gt;not&lt;/strong&gt; run &lt;code&gt;apply&lt;/code&gt; on my real Downloads folder yet — the plan is sitting there for review, which is exactly the workflow I want from a tool that touches my files.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fhn83zz6cf8fykw27sjzz.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fhn83zz6cf8fykw27sjzz.png" alt="Tidy Sloth's generated plan report" width="800" height="714"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'd tell someone reaching for a small model
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Match by position, not by index.&lt;/strong&gt; A model-supplied &lt;code&gt;i&lt;/code&gt; field is another generated token, not a fact.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A correct count proves nothing.&lt;/strong&gt; All twelve rows were there and correctly shaped; the mapping between rows and files was still corrupted.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Validate provenance, not just format.&lt;/strong&gt; Schema validation would have passed every one of those bad renames.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Give the model the smallest job that still helps.&lt;/strong&gt; Here, choosing between 15 folders. Names are mostly mechanical.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The 14 minutes is real, and it's slow — that's a 1B model on a laptop, and most of it is the model, not the rules. A bigger model would name better and cost more energy, which rather misses the point of a sloth.&lt;/p&gt;

&lt;p&gt;Code: &lt;strong&gt;&lt;a href="https://github.com/sank8-2/tidy-sloth" rel="noopener noreferrer"&gt;github.com/sank8-2/tidy-sloth&lt;/a&gt;&lt;/strong&gt; — MIT, one file, no dependencies.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;ollama pull llama3.2:1b
python3 tidysloth.py scan  ~/Downloads   &lt;span class="c"&gt;# read-only summary&lt;/span&gt;
python3 tidysloth.py plan  ~/Downloads   &lt;span class="c"&gt;# proposals + report&lt;/span&gt;
python3 tidysloth.py apply ~/Downloads/.tidysloth/plan.json
python3 tidysloth.py undo  ~/Downloads/.tidysloth/manifest-&lt;span class="k"&gt;*&lt;/span&gt;.json
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Or with no model at all, which still produces a complete and useful plan:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;python3 tidysloth.py plan ~/Downloads &lt;span class="nt"&gt;--no-model&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



</description>
      <category>hf26challenge</category>
      <category>ai</category>
      <category>opensource</category>
      <category>devchallenge</category>
    </item>
    <item>
      <title>I built a listing writer that runs on the seller's own laptop — open-weight models, zero cloud</title>
      <dc:creator>Sanketh Kumar</dc:creator>
      <pubDate>Mon, 19 Oct 2026 05:41:35 +0000</pubDate>
      <link>https://dev.to/sank8/i-built-a-listing-writer-that-runs-on-the-sellers-own-laptop-open-weight-models-zero-cloud-1c27</link>
      <guid>https://dev.to/sank8/i-built-a-listing-writer-that-runs-on-the-sellers-own-laptop-open-weight-models-zero-cloud-1c27</guid>
      <description>&lt;p&gt;&lt;strong&gt;A reseller photographs forty things a week and then writes forty listings by hand.&lt;/strong&gt; That is the tedious part of the side hustle — not the sourcing, not the shipping, the copy. Existing tools fix it by charging per listing and shipping your inventory to somebody else's cloud. I wanted to see how far a &lt;em&gt;local&lt;/em&gt; open-weight model could get instead.&lt;/p&gt;

&lt;p&gt;So I built &lt;strong&gt;Listing Studio&lt;/strong&gt;: describe an item once and it writes a platform-shaped title, description, tags and price range for Etsy, eBay, Depop or Vinted — running entirely on your own machine.&lt;/p&gt;

&lt;p&gt;🔗 &lt;strong&gt;Code:&lt;/strong&gt; &lt;a href="https://github.com/sank8-2/listing-studio" rel="noopener noreferrer"&gt;https://github.com/sank8-2/listing-studio&lt;/a&gt; (MIT)&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fo09occtwzt5s8jxw6xpe.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fo09occtwzt5s8jxw6xpe.png" alt="Listing Studio writing a listing for a vintage wooden birdhouse" width="800" height="647"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Who would actually pay for this
&lt;/h2&gt;

&lt;p&gt;The buyer is a one-person reseller: someone clearing a house, selling vintage finds, or running a small craft business. Their economics are small margins and large volume, and their two real costs with cloud listing tools are:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Per-listing pricing.&lt;/strong&gt; When you list 200 items a month, a subscription or credit pack is a real line item. A local model costs electricity.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Handing over commercially sensitive data.&lt;/strong&gt; Your stock list, your supplier names and your margins are the business. Uploading them to generate ad copy is a strange trade.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Neither of those is a technical argument. They are the reason a local tool exists at all.&lt;/p&gt;

&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;



&lt;pre data-lang="mermaid"&gt;&lt;code&gt;flowchart LR
  A[item notes] --&amp;gt; B[Python: build prompt&amp;lt;br/&amp;gt;+ platform rules]
  B --&amp;gt; C[Ollama&amp;lt;br/&amp;gt;open-weight model]
  C --&amp;gt; D[extract JSON]
  D --&amp;gt; E[Python: clamp title,&amp;lt;br/&amp;gt;normalise tags, price band]
  E --&amp;gt; F[(SQLite history)]
  E --&amp;gt; G[listing + CSV export]&lt;/code&gt;&lt;/pre&gt;



&lt;p&gt;The pipeline has one deliberate rule: &lt;strong&gt;the model writes prose, Python decides every hard limit.&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Etsy: 140-character title, 13 tags × 20 chars&lt;/li&gt;
&lt;li&gt;eBay: 80-character title, no tags&lt;/li&gt;
&lt;li&gt;Depop/Vinted: shorter titles, 5 tags&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The model is never asked to count characters, because small models cannot count. &lt;code&gt;finalise()&lt;/code&gt; clamps titles at a word boundary, lowercases and de-duplicates tags, caps their length, and reorders the price band if the model gets it backwards.&lt;/p&gt;

&lt;p&gt;It is stdlib Python and one HTML file. No pip install, no build step, no account:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;ollama pull llama3.2:1b
python3 app.py        &lt;span class="c"&gt;# → http://localhost:8004&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  What it actually produced
&lt;/h2&gt;

&lt;p&gt;Input:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;item:&lt;/strong&gt; vintage wooden birdhouse · &lt;strong&gt;category:&lt;/strong&gt; garden decor · &lt;strong&gt;materials:&lt;/strong&gt; pine wood · &lt;strong&gt;condition:&lt;/strong&gt; good, some weathering · &lt;strong&gt;details:&lt;/strong&gt; roof lifts off for cleaning; hanging hook on the back; 28 cm tall&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;First run, ~49 seconds on CPU, unedited:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Title:&lt;/strong&gt; Vintage Weathered Pine Wood Birdhouse by Local Carpenter (56/140 chars)&lt;br&gt;
&lt;strong&gt;Tags (12 returned, 13 allowed):&lt;/strong&gt; garden decor · birdhouse · weathered pine wood · wooden birdhouse · hanging hook · pine wood birdhouse · local carpenter · handmade garden deco · garden accessory · wooden bird house · bird house for birds · decorative wooden &lt;em&gt;(clipped to Etsy's 20-char tag limit mid-phrase — you can see the limit doing its work)&lt;/em&gt;&lt;br&gt;
&lt;strong&gt;Price:&lt;/strong&gt; GBP 25–40&lt;br&gt;
&lt;strong&gt;Description:&lt;/strong&gt; &lt;em&gt;the model returned this one empty&lt;/em&gt; — so the deterministic fallback composer wrote it instead. That safety net is not a nice-to-have; it is the difference between an empty listing and a sellable one.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Then a second run on the identical input produced a shorter title (24 chars), only 8 tags, and a &lt;strong&gt;GBP 5–10&lt;/strong&gt; band — that is the screenshot above, and it is the honest headline: &lt;strong&gt;a 1B model is a drafting assistant, not an appraiser.&lt;/strong&gt; Two runs, a 4× spread on price. The seller still sets the price. What the tool removes is the blank page, and it does that at zero marginal cost.&lt;/p&gt;

&lt;h2&gt;
  
  
  The part that did not work
&lt;/h2&gt;

&lt;p&gt;I wanted the photo to be the input — point the camera at the item, get the listing. It is wired up, and it is &lt;strong&gt;not&lt;/strong&gt; the default, for reasons worth sharing:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;On this machine (Ollama 0.18.2, CPU-only) &lt;code&gt;gemma3:1b&lt;/code&gt; and &lt;code&gt;qwen2.5:1.5b&lt;/code&gt; both produced degenerate output — literal token loops like &lt;code&gt;1 1 1 1 1&lt;/code&gt; even at temperature 0. Re-pulling did not help. &lt;code&gt;llama3.2:1b&lt;/code&gt; worked fine, so the runtime is healthy and those blobs are not.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;moondream:latest&lt;/code&gt; described a portrait correctly but returned symbol salad on real product photos, at 130–229 seconds per image.&lt;/li&gt;
&lt;li&gt;With ~4 GB RAM free, holding a vision model and a text model at once made the model server return HTTP 500.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So I shipped text-first and left the photo field in place behind &lt;code&gt;VISION_MODEL&lt;/code&gt;. &lt;strong&gt;The honest version of this project is less magical than the pitch, and it still saves the seller real time.&lt;/strong&gt; I would rather write that down than show a screenshot of a vision pipeline I never got working reliably.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why open matters here specifically
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The privacy claim is structural, not a policy.&lt;/strong&gt; There is no API key and no third-party endpoint; the app talks to &lt;code&gt;127.0.0.1:11434&lt;/code&gt;. You can verify it by pulling the network cable out and watching it keep working.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Model choice stays yours.&lt;/strong&gt; &lt;code&gt;WRITER_MODEL=llama3.1:8b&lt;/code&gt; is the entire upgrade path from "usable draft" to "genuinely good copy". No plan tier, no vendor deciding what you are allowed to run this month.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The price cannot be changed out from under you.&lt;/strong&gt; A tool whose running cost is your own electricity does not get a pricing-page redesign that triples your bill.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For a business tool aimed at people with thin margins, those are not nice-to-haves. They are the product.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/sank8-2/listing-studio
&lt;span class="nb"&gt;cd &lt;/span&gt;listing-studio &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; ollama pull llama3.2:1b &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; python3 app.py
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If you sell things on the side: I would genuinely like to know whether the output survives contact with a real listing, and where the copy falls apart. That feedback is more useful to me than a star.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Built for the Hacktoberfest Open-Source AI Challenge, Week 2 — Side Hustle.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>hf26challenge</category>
      <category>ai</category>
      <category>opensource</category>
      <category>devchallenge</category>
    </item>
    <item>
      <title>Touch Grass: an offline terrace stargazing guide on open-weight Gemma</title>
      <dc:creator>Sanketh Kumar</dc:creator>
      <pubDate>Sun, 11 Oct 2026 06:13:37 +0000</pubDate>
      <link>https://dev.to/sank8/touch-grass-an-offline-terrace-stargazing-guide-on-open-weight-gemma-3i19</link>
      <guid>https://dev.to/sank8/touch-grass-an-offline-terrace-stargazing-guide-on-open-weight-gemma-3i19</guid>
      <description>&lt;p&gt;For the Touch Grass theme I built something that makes the screen the shortest part of the experience: a stargazing guide for Bengaluru terraces that tells you what's actually above the horizon &lt;em&gt;right now&lt;/em&gt; — then sends you outside.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I built
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Terrace Sky Spotter&lt;/strong&gt; — a web app (vanilla HTML/JS + zero-dependency Python server, stdlib only) in two halves:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic astronomy (code, not AI):&lt;/strong&gt; 16 anchor stars from open catalogs (Yale Bright Star Catalogue / HYG, public domain) + real Julian-date → sidereal-time math computes altitude/azimuth for 12.97°N 77.59°E. No GPS, no network. Tap any constellation for its finder hint.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Open-weight narration (local Gemma 3 1B via Ollama):&lt;/strong&gt; a "Narrate my tour" button produces the night's viewing order (setting constellations first), star-hop instructions, myths, and practical tips — plus per-constellation "how do I find it?" answers.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Since the first version I added the three features that make it a real observing tool:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;🌙 &lt;strong&gt;Moon phase + washout filter&lt;/strong&gt; — a synodic model computes illumination (validated against the Mar 3 2026 lunar eclipse); the naked-eye limit slides from ~mag 5.5 to ~mag 2.3 and faint targets get a 🌕 washed-out badge.&lt;/li&gt;
&lt;li&gt;🪐 &lt;strong&gt;Planets&lt;/strong&gt; — low-precision orbital math adds Venus, Mars, Jupiter and Saturn (verified by the Venus-elongation invariant). Saturn at 42° in the east is currently the best thing in the evening sky.&lt;/li&gt;
&lt;li&gt;&lt;p&gt;🕘 &lt;strong&gt;Time picker&lt;/strong&gt; — plan any evening ahead, not just right now.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Live demo:&lt;/strong&gt; run &lt;code&gt;python3 app.py&lt;/code&gt;, open &lt;code&gt;http://localhost:8002&lt;/code&gt; (needs Ollama + &lt;code&gt;gemma3:1b&lt;/code&gt;; tour narrates in ~30s on CPU)&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Code:&lt;/strong&gt; &lt;a href="https://github.com/sank8-2/terrace-sky-spotter" rel="noopener noreferrer"&gt;https://github.com/sank8-2/terrace-sky-spotter&lt;/a&gt; (MIT)&lt;/p&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Who it's for
&lt;/h2&gt;

&lt;p&gt;My roommate Nagaraj and me — terrace sessions in Bengaluru, phone brightness down, eyes up. I took it outside on a bright moonlit night and gazed for real: the moonlight washed out the faint stuff, exactly as the app's washout filter now predicts — the Pegasus square, Vega and Altair punched through while Pisces vanished. That field night is literally what the 🌕 badge is built from.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why open matters here
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Works with no signal.&lt;/strong&gt; The catalog is bundled data and the model runs on-device. A closed-API guide dies the moment the bars disappear — exactly where you need it most: terraces, trails, villages.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Location stays yours.&lt;/strong&gt; Stargazing needs your coordinates; here they never leave the laptop.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Free + swappable.&lt;/strong&gt; One env var changes the model tomorrow.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  The honest part: don't let the model invent the sky
&lt;/h2&gt;

&lt;p&gt;Testing caught the key failure mode: asked about Cassiopeia while it was below the horizon, the model cheerfully described it as visible "high up". So the app never lets Gemma state positions — the deterministic math owns &lt;em&gt;what's up and where&lt;/em&gt;, and the model only narrates &lt;em&gt;how to find&lt;/em&gt; what code already verified. Below-horizon queries get an explicit "not up right now, here's when to catch it" path. Same architecture as my weekend entry's allergy guardrail: code enforces the safety property, the model supplies the words.&lt;/p&gt;

&lt;p&gt;Also entering &lt;strong&gt;Best Use of Gemma&lt;/strong&gt; — narrated end-to-end by open-weight &lt;code&gt;gemma3:1b&lt;/code&gt;, running locally.&lt;/p&gt;

</description>
      <category>hf26challenge</category>
      <category>ai</category>
      <category>opensource</category>
      <category>devchallenge</category>
    </item>
    <item>
      <title>Build for a Friend: a local-first meal planner on open-weight Gemma</title>
      <dc:creator>Sanketh Kumar</dc:creator>
      <pubDate>Mon, 05 Oct 2026 03:33:46 +0000</pubDate>
      <link>https://dev.to/sank8/build-for-a-friend-a-local-first-meal-planner-on-open-weight-gemma-59po</link>
      <guid>https://dev.to/sank8/build-for-a-friend-a-local-first-meal-planner-on-open-weight-gemma-59po</guid>
      <description>&lt;p&gt;My roommate Nagaraj needed high-protein vegetarian food on a ₹2000/week budget — while strictly avoiding peanuts and lactose. So for the Build for a Friend theme, I built him a meal planner that runs &lt;strong&gt;entirely on his laptop&lt;/strong&gt;: no accounts, no API keys, no cloud.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I built
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Friend Meal Planner&lt;/strong&gt; — a web app (vanilla HTML/JS + a zero-dependency Python server, stdlib only) that generates a 7-day Indian meal plan with a grocery list and budget check. You fill in your friend's profile — diet, allergies, dislikes, budget — and the plan streams in, token by token.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fcc3dcjujxb823726qd5g.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fcc3dcjujxb823726qd5g.png" alt="Meal Planner" width="789" height="994"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Live demo:&lt;/strong&gt; run &lt;code&gt;python3 app.py&lt;/code&gt; and open &lt;code&gt;http://localhost:8000&lt;/code&gt; (needs Ollama + &lt;code&gt;gemma3:1b&lt;/code&gt; pulled; ~40s per plan on CPU)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Code:&lt;/strong&gt; &lt;a href="https://github.com/sank8-2/friend-meal-planner" rel="noopener noreferrer"&gt;https://github.com/sank8-2/friend-meal-planner&lt;/a&gt; (MIT)&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Who it's for
&lt;/h2&gt;

&lt;p&gt;Nagaraj, my roommate in Bengaluru. Hostel-style cooking, 30-minute recipes, ingredients from local markets, everything priced in ₹. I handed him the first plan — his verdict: &lt;em&gt;"now that's cooking"&lt;/em&gt; (he also immediately asked for a recipe mode and non-veg mode, so that's next).&lt;/p&gt;

&lt;h2&gt;
  
  
  Why open matters here
&lt;/h2&gt;

&lt;p&gt;Open-weight AI is what makes this project &lt;em&gt;possible&lt;/em&gt;, not just cheaper:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Health data stays home.&lt;/strong&gt; Allergies and eating habits are sensitive. With a local open-weight model, none of it touches a server anyone else controls — a closed API would mean shipping my friend's health profile to a third party on every generation.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Works offline.&lt;/strong&gt; Hostel Wi-Fi dies constantly; the planner doesn't care.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Free forever.&lt;/strong&gt; No per-token billing against a student budget.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Swappable.&lt;/strong&gt; Gemma 3 1B today, something bigger tomorrow via one env var — no vendor lock-in.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  The honest part: small models slip, so I built a guardrail
&lt;/h2&gt;

&lt;p&gt;Testing caught something important: the 1B model kept suggesting paneer and yogurt to my &lt;em&gt;lactose-allergic&lt;/em&gt; friend, despite explicit instructions. Rather than hide that, I added a &lt;strong&gt;deterministic allergy guardrail&lt;/strong&gt; — every plan is scanned line-by-line against an allergen keyword map (lactose → milk, curd, paneer, ghee…) and an ⚠️ Allergy Check section is appended flagging risky dishes for review. Prompt engineering sets the intent; code enforces the safety property. That's the architecture I'd defend: never let a probabilistic model be the last word on food safety.&lt;/p&gt;

&lt;p&gt;This also enters the &lt;strong&gt;Best Use of Gemma&lt;/strong&gt; category — the whole app is Gemma-powered, running the open-weight &lt;code&gt;gemma3:1b&lt;/code&gt; model locally through Ollama.&lt;/p&gt;

</description>
      <category>hf26challenge</category>
      <category>weekendchallenge</category>
      <category>devchallenge</category>
      <category>opensource</category>
    </item>
    <item>
      <title>GoFr!?</title>
      <dc:creator>Sanketh Kumar</dc:creator>
      <pubDate>Wed, 23 Oct 2024 05:06:09 +0000</pubDate>
      <link>https://dev.to/sank8/gofr-2h4m</link>
      <guid>https://dev.to/sank8/gofr-2h4m</guid>
      <description>&lt;p&gt;Consider GoFr to be your own microservice helper. This framework, which is based on the robust Go language, handles the menial duties so you can concentrate on creating incredible applications.&lt;/p&gt;

&lt;p&gt;Important attributes:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;With built-in middleware, there's no need to start from scratch. GoFr saves you time and effort by including necessary middlewares like authentication, error handling, and logging.&lt;/li&gt;
&lt;li&gt;Support for Datastores, File Systems, and Pub/Sub: Easily connect to your preferred file systems, messaging apps, and databases. You're insured by GoFr.
Integrated Monitoring: Use the integrated tracing, monitoring, and logging features of GoFr to closely monitor your microservices.&lt;/li&gt;
&lt;li&gt;Opinionated Approach: GoFr adheres to a precise and dependable framework that directs you toward best practices and steers clear of typical errors.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Why Opt for GoFr?&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Quicker Development: Focus more on creating amazing features and less on boilerplate.&lt;/li&gt;
&lt;li&gt;Enhanced Performance: GoFr creates high-performance microservices by utilizing Go's capabilities.&lt;/li&gt;
&lt;li&gt;Enhanced Reliability: GoFr assists you in developing dependable and strong applications with integrated monitoring and error management.&lt;/li&gt;
&lt;li&gt;Simplified Deployment: GoFr's Kubernetes design simplifies the deployment and management of your microservices.&lt;/li&gt;
&lt;li&gt;Are you prepared to accelerate the development of microservices? &lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://gofr.dev/" rel="noopener noreferrer"&gt;Try GoFr&lt;/a&gt; and see the difference for yourself.&lt;/p&gt;

</description>
      <category>go</category>
      <category>gofr</category>
    </item>
  </channel>
</rss>
