<?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: Mario Vargas</title>
    <description>The latest articles on DEV Community by Mario Vargas (@drakenmario).</description>
    <link>https://dev.to/drakenmario</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%2F452294%2F633991a9-107c-4693-b58c-ec4d6288d2f3.jpg</url>
      <title>DEV Community: Mario Vargas</title>
      <link>https://dev.to/drakenmario</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/drakenmario"/>
    <language>en</language>
    <item>
      <title>I used Pi subagents to test my hackathon project against real code changes</title>
      <dc:creator>Mario Vargas</dc:creator>
      <pubDate>Wed, 18 Nov 2026 22:48:12 +0000</pubDate>
      <link>https://dev.to/drakenmario/i-used-pi-subagents-to-test-my-hackathon-project-against-real-code-changes-pll</link>
      <guid>https://dev.to/drakenmario/i-used-pi-subagents-to-test-my-hackathon-project-against-real-code-changes-pll</guid>
      <description>&lt;p&gt;I asked a Pi agent to test LessCode against real pull requests. It sent one subagent to find TypeScript repositories and candidate PRs, then assigned the test runs to other subagents.&lt;/p&gt;

&lt;p&gt;That was the moment I thought, &lt;em&gt;Okay, this is doable.&lt;/em&gt; Finding useful test cases and running the tool against each one was a job I would normally do by hand. Now I could give that job to a workflow I’d built and see what came back from each run.&lt;/p&gt;

&lt;h2&gt;
  
  
  One agent found the cases. Others ran them.
&lt;/h2&gt;

&lt;p&gt;I built LessCode for the Meta Global AI Developer Hackathon using Pi and Pi subagents. LessCode reviews JavaScript and TypeScript changes. It gives you a review that points to the source and can offer a patch when a supported edit passes its checks. For the hackathon, I used Meta’s Muse Spark Open through Baseten to propose findings and edits.&lt;/p&gt;

&lt;p&gt;I call my Pi setup my &lt;em&gt;software factory&lt;/em&gt;. Here’s what it did for this evaluation:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Role&lt;/th&gt;
&lt;th&gt;Job&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Pi coordinator&lt;/td&gt;
&lt;td&gt;Organize the evaluation and assign the work.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sourcing subagent&lt;/td&gt;
&lt;td&gt;Find TypeScript repositories and candidate PRs.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Testing subagents&lt;/td&gt;
&lt;td&gt;Run LessCode on their assigned changes and report what happened.&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  The evaluation got its own VM
&lt;/h2&gt;

&lt;p&gt;My home lab runs Proxmox, so I created a VM just for this evaluation. The coordinator and its subagents worked inside that VM, apart from the rest of my lab.&lt;/p&gt;

&lt;p&gt;The VM gave the whole job a place to run. Inside it, the assignments gave each runner a narrower job. I wanted to know which result came from which change, so I could look at each case on its own.&lt;/p&gt;

&lt;h2&gt;
  
  
  What one run taught me
&lt;/h2&gt;

&lt;p&gt;One case I documented was a &lt;strong&gt;JavaScript refactor in &lt;code&gt;is-number&lt;/code&gt;&lt;/strong&gt;. I ran the same pinned change through LessCode’s engine, CLI, and MCP interface. All three finished without proposing a patch.&lt;/p&gt;

&lt;p&gt;Then I read the reviews. Two of them said an expression using &lt;code&gt;typeof num&lt;/code&gt; would throw a &lt;code&gt;ReferenceError&lt;/code&gt;. I checked the functions in isolation. For the inputs I tested, they returned &lt;code&gt;false&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;That gave me a specific thing to improve. I added checks that let LessCode verify certain claims against evidence from the code it reviewed, then ran the same change again. This time, the CLI and MCP reports correctly described the behavior for those inputs. The CLI also withheld two claims it couldn’t support with citations.&lt;/p&gt;

&lt;h2&gt;
  
  
  Watch the hackathon demo
&lt;/h2&gt;

&lt;p&gt;Here’s the three-minute LessCode demo I made for the Meta Global AI Developer Hackathon.&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/L2IOotYcpso" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;h2&gt;
  
  
  What I’ll reuse
&lt;/h2&gt;

&lt;p&gt;The part I want to keep is the ability to go back to the same case. One agent can find a change worth reviewing, another can run the tool, and I can take a claim from the output and check it myself. After I change the tool, I can rerun the exact revision and see whether that particular result improved.&lt;/p&gt;

&lt;p&gt;I built a code reviewer at the hackathon. I came away with a way to test the reviewer that I want to use on my next project, too.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;If you use subagents for testing, what do you save so you can rerun a case and compare the results later?&lt;/strong&gt;&lt;/p&gt;

</description>
      <category>showdev</category>
      <category>ai</category>
      <category>typescript</category>
      <category>tooling</category>
    </item>
    <item>
      <title>My local LLM suggested 27 references. Only 7 passed verification.</title>
      <dc:creator>Mario Vargas</dc:creator>
      <pubDate>Wed, 18 Nov 2026 22:06:56 +0000</pubDate>
      <link>https://dev.to/drakenmario/my-local-llm-suggested-27-references-only-7-passed-verification-4nm9</link>
      <guid>https://dev.to/drakenmario/my-local-llm-suggested-27-references-only-7-passed-verification-4nm9</guid>
      <description>&lt;p&gt;I record a lot of lectures and meetings, but I almost never watch them again.&lt;/p&gt;

&lt;p&gt;What I actually want afterward is something I can read: a summary of what was discussed, the main topics, and any books or papers that were mentioned.&lt;/p&gt;

&lt;p&gt;So I started building &lt;a href="https://github.com/Mariotutal/video-to-doc" rel="noopener noreferrer"&gt;vtd (video-to-doc)&lt;/a&gt;, an open-source CLI that turns audio and video recordings into structured Markdown or PDF documents.&lt;/p&gt;

&lt;p&gt;At some point, I thought it would be useful to have the AI identify books and academic papers mentioned in a recording.&lt;/p&gt;

&lt;p&gt;That sounded like a fairly straightforward feature.&lt;/p&gt;

&lt;p&gt;Then I tested it with a real two-hour lecture.&lt;/p&gt;

&lt;p&gt;My local model (&lt;code&gt;qwen3.6:35b&lt;/code&gt; running through Ollama) suggested &lt;strong&gt;27 references. I could only verify 7 of them&lt;/strong&gt; against Open Library and Crossref.&lt;/p&gt;

&lt;p&gt;That test led me to add a catalog verification step before a reference can appear in the final document.&lt;/p&gt;

&lt;h2&gt;
  
  
  What vtd does
&lt;/h2&gt;

&lt;p&gt;The basic idea is simple. Give it a recording:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;vtd lecture.mp4
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Without enabling AI, vtd uses &lt;code&gt;whisper.cpp&lt;/code&gt; to transcribe the recording, splits it into sections based on pauses, extracts frames from the video, and generates a document combining everything.&lt;/p&gt;

&lt;p&gt;It works locally, without an API key.&lt;/p&gt;

&lt;p&gt;If you add &lt;code&gt;--with-ai&lt;/code&gt;, a language model generates notes that include:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Summaries for each section&lt;/li&gt;
&lt;li&gt;Main topics and themes&lt;/li&gt;
&lt;li&gt;Follow-up tasks&lt;/li&gt;
&lt;li&gt;A glossary of relevant terms&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;You can use a hosted model or a local one through Ollama.&lt;/p&gt;

&lt;p&gt;I use Ollama to run the model locally.&lt;/p&gt;

&lt;h2&gt;
  
  
  Adding a verification step
&lt;/h2&gt;

&lt;p&gt;I added a &lt;code&gt;--references&lt;/code&gt; flag.&lt;/p&gt;

&lt;p&gt;When enabled, the model identifies two types of works:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;References it believes were mentioned in the transcript.&lt;/li&gt;
&lt;li&gt;Additional works that might be useful as further reading.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;For each one, it proposes an author, title, and publication year.&lt;/p&gt;

&lt;p&gt;But instead of trusting that information, vtd searches two public catalogs:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Open Library&lt;/strong&gt;, primarily for books.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Crossref&lt;/strong&gt;, for academic publications and other indexed works.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;To accept a reference, vtd checks that:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The title matches word for word, allowing spelling variations or a missing subtitle.&lt;/li&gt;
&lt;li&gt;The first author's surname matches, including minor spelling or transliteration differences.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For example, &lt;code&gt;Vygotsky&lt;/code&gt; and &lt;code&gt;Vygotskiĭ&lt;/code&gt; can still match.&lt;/p&gt;

&lt;p&gt;The final citation uses &lt;strong&gt;the catalog’s metadata&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;To organize the bibliography, vtd checks the relevant transcript section for the author’s capitalized surname or a title of at least four words. It lists verified works with transcript evidence under &lt;strong&gt;Works referred to in the recording&lt;/strong&gt; and other verified suggestions under &lt;strong&gt;Further reading&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Testing it with a public recording
&lt;/h2&gt;

&lt;p&gt;I added a reproducible example to the repository.&lt;/p&gt;

&lt;p&gt;I used President Kennedy's 1962 speech at Rice University, an 18-minute public-domain recording.&lt;/p&gt;

&lt;p&gt;Here's the command:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;vtd rice-speech.ogg \
  --language en \
  --whisper-model large-v3-turbo \
  --with-ai \
  --provider ollama \
  --ai-model qwen3.6:35b \
  --content-type lecture \
  --references \
  --pdf
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;I ran it on a Mac with an M4 Pro and 48 GB of memory.&lt;/p&gt;

&lt;p&gt;The results:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Step&lt;/th&gt;
&lt;th&gt;Time&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Processing without AI&lt;/td&gt;
&lt;td&gt;36 seconds&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;AI notes, references, and PDF&lt;/td&gt;
&lt;td&gt;101 additional seconds&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The Whisper model was already downloaded.&lt;/p&gt;

&lt;p&gt;These timings are from a single run.&lt;/p&gt;

&lt;h3&gt;
  
  
  Bibliography
&lt;/h3&gt;

&lt;p&gt;One of the eight proposed works matched a catalog record and appeared in the bibliography as further reading.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Works referred to in the recording&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;None detected.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Further reading (not mentioned in the recording)&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;H. Kahn (1960). &lt;em&gt;On thermonuclear war&lt;/em&gt;. &lt;a href="https://openlibrary.org/works/OL9984394W" rel="noopener noreferrer"&gt;Open Library&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;The project is open source and MIT-licensed:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;GitHub:&lt;/strong&gt; &lt;a href="https://github.com/Mariotutal/video-to-doc" rel="noopener noreferrer"&gt;https://github.com/Mariotutal/video-to-doc&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;If you work with recorded lectures, meetings, or presentations, feel free to give it a try.&lt;/p&gt;

&lt;p&gt;I’d be interested to hear how you use it with your own recordings. Share your experience in the comments or open an issue on GitHub.&lt;/p&gt;

</description>
      <category>showdev</category>
      <category>ai</category>
      <category>opensource</category>
      <category>llm</category>
    </item>
  </channel>
</rss>
