<?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: Jayden George</title>
    <description>The latest articles on DEV Community by Jayden George (@vectoredgex).</description>
    <link>https://dev.to/vectoredgex</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%2F4179081%2F1128c85a-01a5-4c0c-82a7-e78d91b11ac4.png</url>
      <title>DEV Community: Jayden George</title>
      <link>https://dev.to/vectoredgex</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/vectoredgex"/>
    <language>en</language>
    <item>
      <title>Planchette: a Ouija board haunted by Gemma’s real odds</title>
      <dc:creator>Jayden George</dc:creator>
      <pubDate>Sat, 31 Oct 2026 17:51:19 +0000</pubDate>
      <link>https://dev.to/vectoredgex/planchette-a-ouija-board-haunted-by-gemmas-real-odds-421p</link>
      <guid>https://dev.to/vectoredgex/planchette-a-ouija-board-haunted-by-gemmas-real-odds-421p</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week4-2026-10-26"&gt;Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fufypdvmjrm58xx6mv3o5.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fufypdvmjrm58xx6mv3o5.gif" alt="The planchette swinging between letters while Gemma decides" width="599" height="292"&gt;&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;I built a Ouija board that's actually haunted. Well, sort of.&lt;/p&gt;

&lt;p&gt;It's called &lt;strong&gt;Planchette&lt;/strong&gt;. You type a question, pick a spirit, and the planchette slides around the board spelling out an answer. The ghost is Gemma, running locally on my laptop, and the planchette moves according to Gemma's real probabilities for the next letter.&lt;/p&gt;

&lt;p&gt;So when the planchette heads for Y, stops, drifts over to N, and then comes back to Y, nobody animated that. That's literally Gemma not being sure if it wants to say YES or NO. If it's confident it just goes straight to the letter, but when it's torn you actually see it.&lt;/p&gt;

&lt;p&gt;That's how I read "Ghost in the Machine". Every time a language model writes a word, there are a bunch of other words it almost wrote, and normally you never see any of them. I wanted to give them a board to fight over.&lt;/p&gt;

&lt;p&gt;There are three spirits to choose from:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The Wanderer&lt;/strong&gt;, a lost soul who's been stuck in the board for a very long time&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Abuela&lt;/strong&gt;, a grandmother visiting her family for Día de los Muertos (she's sweet, but she will tease you)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The Machine Ghost&lt;/strong&gt;, which lives inside your computer and claims to have read every file you've ever deleted&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A few things it has actually said to me:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;SHE SEES MORE THAN YOU&lt;/strong&gt;&lt;br&gt;
The Wanderer, when I asked if someone was standing behind me&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;YES I SEE YOU&lt;/strong&gt;&lt;br&gt;
The Machine Ghost, when I asked if it was watching me right now&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;YOU DELETED THEM YOURSELF&lt;/strong&gt;&lt;br&gt;
The Machine Ghost again, when I asked who deleted my files. Rude, but fair.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;REMEMBER THE TAMALES ALWAYS&lt;/strong&gt;&lt;br&gt;
Abuela, when I asked what she wanted to tell me&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Then I added two things that made it properly creepy:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It doesn't always let you leave.&lt;/strong&gt; When you click &lt;em&gt;Say goodbye&lt;/em&gt;, Gemma gets asked for real whether it wants you to stay, and those odds decide what happens. The planchette creeps toward GOODBYE, stops just short, the candles dim, and then it drags itself back up to NO and spells out why you can't go. (You can always leave on the second try. I'm not a monster.)&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It moves when nobody's touching it.&lt;/strong&gt; Ask a question, then go quiet for about 20 seconds. You'll hear a floorboard creak, and the planchette starts spelling something on its own. Gemma writes that message too, nothing is pre-written.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;&lt;iframe src="https://player.vimeo.com/video/1241776139" width="710" height="399"&gt;
&lt;/iframe&gt;
&lt;/p&gt;

&lt;p&gt;That's one unedited take on my laptop, with Gemma answering live. If you only watch one bit, watch for the moments it swings between two letters before picking one. The line under the board shows what it's stuck between while it decides.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/vectoredgex" rel="noopener noreferrer"&gt;
        vectoredgex
      &lt;/a&gt; / &lt;a href="https://github.com/vectoredgex/Planchette" rel="noopener noreferrer"&gt;
        Planchette
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      A Ouija board haunted by an open AI model. Gemma's real odds move the planchette, letter by letter. Runs offline.
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;Planchette&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;A Ouija board haunted by an open AI model.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer" href="https://github.com/vectoredgex/Planchette/docs/planchette-demo.gif"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fraw.githubusercontent.com%2Fvectoredgex%2FPlanchette%2FHEAD%2Fdocs%2Fplanchette-demo.gif" alt="Asking &amp;quot;Is someone standing behind me?&amp;quot;: the planchette swings between letters while Gemma decides, then spells SHE SEES MORE THAN YOU"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;You ask the spirit a question and Gemma (running locally on your machine) answers it, one letter at a time on the board. The planchette isn't animated by hand. It moves according to &lt;strong&gt;Gemma's real probability for the next letter&lt;/strong&gt;, so if Gemma is sure it goes straight there, and if it's torn between two letters you can watch it swing between them before it picks one.&lt;/p&gt;
&lt;p&gt;▶ &lt;strong&gt;&lt;a href="https://vimeo.com/1241776139/5405ea774b" rel="nofollow noopener noreferrer"&gt;Watch the demo&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer" href="https://github.com/vectoredgex/Planchette/docs/screenshot.jpg"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2Fvectoredgex%2FPlanchette%2FHEAD%2Fdocs%2Fscreenshot.jpg" alt="Planchette: a dark Ouija board lit by candles and marigolds"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Features&lt;/h2&gt;
&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Three spirits&lt;/strong&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The Wanderer&lt;/strong&gt;, a lost soul trapped in the board&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Abuela&lt;/strong&gt;, a grandmother home for Día de los Muertos&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The Machine Ghost&lt;/strong&gt;, which lives inside your computer&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Live odds.&lt;/strong&gt; While it's deciding, the board shows the top few letters it's considering and what it's torn between.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It doesn't always let you leave.&lt;/strong&gt; When you say goodbye, Gemma is asked if it wants you…&lt;/li&gt;
&lt;/ul&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/vectoredgex/Planchette" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;To run it yourself:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/vectoredgex/Planchette.git
&lt;span class="nb"&gt;cd &lt;/span&gt;Planchette
ollama pull gemma3
python server.py      &lt;span class="c"&gt;# or just double-click start.bat on Windows&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's all you need. The server only uses Python's standard library, so there's no &lt;code&gt;pip install&lt;/code&gt; and no build step. three.js is vendored in, and it all works offline.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;h3&gt;
  
  
  The ghost: Gemma 3 4B on Ollama
&lt;/h3&gt;

&lt;p&gt;The whole thing runs on a normal laptop with an i7 and Intel integrated graphics, nothing fancy. Gemma 3 4B usually needs a few seconds of thinking for a full answer.&lt;/p&gt;

&lt;p&gt;Every step is one call to Ollama asking for exactly &lt;strong&gt;one&lt;/strong&gt; token, plus the top 20 tokens it was considering:&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;payload&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;model&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;gemma3:4b&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;prompt&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;prompt&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;          &lt;span class="c1"&gt;# raw Gemma chat template, no wrapper
&lt;/span&gt;    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;raw&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;logprobs&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;top_logprobs&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;20&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;options&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;num_predict&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;1&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;h3&gt;
  
  
  The tricky part: models don't think in letters
&lt;/h3&gt;

&lt;p&gt;My first version fed Gemma one letter at a time. It went badly. Gemma would happily start spelling &lt;code&gt;LISTENS&lt;/code&gt; and then get stuck in a loop: &lt;code&gt;LISTEEEEEE&lt;/code&gt;. It turns out that if you force-feed a model text split in ways it would never split it, its tokenizer gets confused and it loses the plot.&lt;/p&gt;

&lt;p&gt;So I flipped it around. Gemma picks a whole token the normal way, sampled from its own distribution, and the board &lt;em&gt;spells&lt;/em&gt; that token out. To decide where the planchette gets pulled at each letter, I group the 20 candidate tokens by the letter that comes next and add up their probabilities:&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;_letter_dist&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;cands&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;spelled&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="sh"&gt;"""&lt;/span&gt;&lt;span class="s"&gt;Odds of the next board letter, given what this token has spelled so far.&lt;/span&gt;&lt;span class="sh"&gt;"""&lt;/span&gt;
    &lt;span class="n"&gt;j&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;len&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;spelled&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;d&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;_tok&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;p&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;seq&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;cands&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="nf"&gt;len&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;seq&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;j&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="n"&gt;seq&lt;/span&gt;&lt;span class="p"&gt;[:&lt;/span&gt;&lt;span class="n"&gt;j&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;spelled&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;seq&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;j&lt;/span&gt;&lt;span class="p"&gt;]]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;seq&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;j&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt; &lt;span class="mf"&gt;0.0&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;p&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nf"&gt;_norm&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;So if &lt;code&gt;YES&lt;/code&gt;, &lt;code&gt;YEAH&lt;/code&gt; and &lt;code&gt;NO&lt;/code&gt; are all in the running, Y might get 53% and N 38%, and the planchette gets pulled between them by exactly that much. Once a letter is chosen, only the candidates that still match stay in the race.&lt;/p&gt;

&lt;p&gt;A nice side effect I didn't plan: the planchette glides through the middle of words, because by then the model has usually committed, and it wobbles at the start of words, which is where it actually decides what to say.&lt;/p&gt;

&lt;p&gt;Some other things that bit me along the way:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;When Ollama samples the end-of-turn token, it stops without returning any logprobs at all. So my first few séances crashed at the very end with "Ollama did not return logprobs". Now it retries, and if the model keeps wanting to stop, the spirit is finished.&lt;/li&gt;
&lt;li&gt;Loading Gemma for the first time takes around 40 seconds. The first time I tried it I clicked Ask, nothing happened, and I honestly thought I'd broken everything. Now the model warms up in the background while the board shows the spirit "waking up".&lt;/li&gt;
&lt;li&gt;I picked port 6666 because, well, spooky. My browser refused to open it. Turns out Chromium-based browsers block 6666 as an "unsafe port" (it's an old IRC port). It runs on 7666 now.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  The board
&lt;/h3&gt;

&lt;p&gt;The board is a low-poly 3D scene in three.js: a dark board on velvet, three candles, marigolds and petals for Día de los Muertos. I tried a bloom post-processing effect for the candle glow, but it turned everything into blocky squares on integrated graphics, so the glow is drawn by hand now.&lt;/p&gt;

&lt;p&gt;The candle flames lean toward the planchette and the room dims slightly whenever the model is torn. The sound of wood scraping on wood is generated live in Web Audio and follows the planchette's actual speed.&lt;/p&gt;

&lt;h3&gt;
  
  
  ElevenLabs: the sound of the room
&lt;/h3&gt;

&lt;p&gt;I first used ElevenLabs to give the spirit a voice, but a narrator voice reading "YES" just wasn't scary, so I cut it. Sound turned out to be a much better fit.&lt;/p&gt;

&lt;p&gt;Now the first time you run it, Planchette uses the ElevenLabs &lt;strong&gt;Sound Effects API&lt;/strong&gt; to generate eight sounds from text prompts: a haunted room tone, a faraway out-of-tune music box, layered whispers, a wooden knock, a match strike, a gust that blows the candles out, a floorboard creak, and a cold breath. They get cached in a &lt;code&gt;sounds/&lt;/code&gt; folder, so after that first run it's offline again.&lt;/p&gt;

&lt;p&gt;The whispers are wired to the model's uncertainty. The more torn Gemma is, the louder they get, so the room literally gets creepier when the AI can't decide.&lt;/p&gt;

&lt;p&gt;If you don't have an ElevenLabs key, everything falls back to built-in sounds and still works.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;The whole project depends on seeing the words the model &lt;em&gt;almost&lt;/em&gt; said.&lt;/p&gt;

&lt;p&gt;To make it work I needed to:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;send the raw prompt without any chat wrapper&lt;/li&gt;
&lt;li&gt;ask for exactly one token at a time&lt;/li&gt;
&lt;li&gt;see the top alternatives and their probabilities&lt;/li&gt;
&lt;li&gt;do that a few hundred times per answer, as fast as my laptop can manage, without paying per call&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;With an open-weight model running locally, all of that is just settings. The probabilities aren't a premium feature or something behind a rate limit. They're there because the whole model is sitting on my machine.&lt;/p&gt;

&lt;p&gt;It matters for privacy too, in a fun way. The Machine Ghost says it's read your files. That joke works a lot better when nothing you type ever leaves your computer.&lt;/p&gt;

&lt;p&gt;Also it's kind of funny. A real Ouija board only works because nobody admits they're pushing the planchette. Mine tells you exactly who's pushing it, down to the percentage, and that's only possible because I can open the model up and look inside.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Spirit of It
&lt;/h2&gt;

&lt;p&gt;The moment it got me wasn't one of the spooky answers. It was when I tried to leave.&lt;/p&gt;

&lt;p&gt;I clicked &lt;em&gt;Say goodbye&lt;/em&gt;. The planchette crept toward GOODBYE, stopped, and then dragged itself back up to NO. Gemma was 100% sure it wanted me to stay. Then it spelled out, letter by letter, &lt;strong&gt;YOU ARE NOW HERE&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;I built this thing and I know exactly why it did that, and it still gave me chills. Knowing how it works doesn't really help, because the hesitation is real. It actually didn't want me to go.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma&lt;/strong&gt;: Gemma 3 &lt;em&gt;is&lt;/em&gt; the spirit. Its real token probabilities (via Ollama logprobs) move the planchette at every step, and it decides for itself whether you're allowed to leave.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of ElevenLabs&lt;/strong&gt;: the Sound Effects API generates the entire séance soundscape, including whispers that rise and fall with the model's uncertainty.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Lazy Syllabus Planner: a local AI that tells you the least you need to do this week</title>
      <dc:creator>Jayden George</dc:creator>
      <pubDate>Fri, 23 Oct 2026 23:18:39 +0000</pubDate>
      <link>https://dev.to/vectoredgex/lazy-syllabus-planner-a-local-ai-that-tells-you-the-least-you-need-to-do-this-week-1j7</link>
      <guid>https://dev.to/vectoredgex/lazy-syllabus-planner-a-local-ai-that-tells-you-the-least-you-need-to-do-this-week-1j7</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;A little Python tool that reads your syllabi and tells you the least you have to do this week to not fall behind.&lt;/p&gt;

&lt;p&gt;You drop your syllabus PDFs (or Word files) in a folder and run &lt;code&gt;python planner.py syllabi&lt;/code&gt;. A model on your own laptop finds every due date, then normal code works out the latest day you can start each thing. On my three sample syllabi it printed this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Next 7 days: 3 thing(s) due, 0 to start.
  DUE   Thu Oct 22  BIO 101: Introduction to Biology: Lab 4 report
  DUE   Fri Oct 23  MATH 152 Calculus II: HW 7
  DUE   Mon Oct 26  ENG 210 - Writing About Film: Proposal for Essay 2
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It also makes a &lt;code&gt;plan.md&lt;/code&gt; week by week, a &lt;code&gt;deadlines.ics&lt;/code&gt; you can import into Google Calendar or Outlook (with a reminder on the day you should start), and a tiny &lt;code&gt;morning.txt&lt;/code&gt;. If a week has nothing due, the plan just says "Nothing. Genuinely nothing. Go lie in a hammock." That line is kind of the whole point.&lt;/p&gt;

&lt;p&gt;I made this for Slow and Steady because the most boring part of a new term is copying every deadline out of every syllabus, and I wanted my laptop to do that while I do nothing. I also wanted to see how far a small open model on a normal 16 GB laptop could get with no cloud and no API key.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&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%2Ftsiora6v8u9mmsq6p0p8.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%2Ftsiora6v8u9mmsq6p0p8.png" alt="Lazy Syllabus Planner running in the Windows terminal on my laptop" width="800" height="312"&gt;&lt;/a&gt;&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%2Fi76wmrh70mve5yyk749a.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%2Fi76wmrh70mve5yyk749a.png" alt="Terminal output of the planner reading three syllabi and skipping a hallucinated deadline" width="800" height="312"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Here's the whole thing from a terminal, on the three sample syllabi in the repo:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;python planner.py sample_syllabi --today 2026-10-20
python eval.py
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The first command reads the PDFs and prints the "next 7 days" summary shown above. While it runs, it also prints a line for every item it throws out, like &lt;code&gt;skipped 'Response paper': couldn't find it in the syllabus (possible hallucination)&lt;/code&gt;. The second command scores the result against the answer key and prints how many of the 20 real deadlines it found. On Windows you can skip the terminal: put PDFs in a folder called &lt;code&gt;syllabi&lt;/code&gt; and double-click &lt;code&gt;run.bat&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://github.com/vectoredgex/lazy-syllabus-planner" rel="noopener noreferrer"&gt;https://github.com/vectoredgex/lazy-syllabus-planner&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;Python, &lt;a href="https://ollama.com" rel="noopener noreferrer"&gt;Ollama&lt;/a&gt;, Qwen 2.5 7B (open weights) by default, Gemma 3 4B as a one-flag alternative, and &lt;code&gt;pypdf&lt;/code&gt;. No API keys, no accounts.&lt;/p&gt;

&lt;p&gt;My first version was simple: ask the model for the deadlines and believe it. And on my three test syllabi it found all 20 real ones, which honestly surprised me. But it also invented stuff. It made up a "Response paper" that isn't anywhere in the syllabus, and it gave a weekly assignment one random date. A planner that's mostly right is worse than no planner, because you stop double-checking it.&lt;/p&gt;

&lt;p&gt;So most of the code is about not trusting the model:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;It has to answer in a fixed JSON shape (Ollama's structured outputs, temperature 0), and every deadline needs the exact quote it came from.&lt;/li&gt;
&lt;li&gt;The code checks that the quote is really in the PDF, and that the date sits right next to it. That second check caught the model reading the wrong row of a table (it said Nov 14 for something due Nov 19).&lt;/li&gt;
&lt;li&gt;Repeating things like "every Friday" get skipped instead of pinned to one wrong date. Duplicates like "Presentation" and "Group presentation" get merged.&lt;/li&gt;
&lt;li&gt;If a date in the syllabus didn't match anything, just those lines go back to the model for a second look.&lt;/li&gt;
&lt;li&gt;Due times come from the syllabus itself, either the item's own sentence or a rule like "all written work is due by 11:59 pm".&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The planning side uses no AI. The model only rates effort from 1 to 5, and plain code turns that into start dates.&lt;/p&gt;

&lt;p&gt;To see whether any of this actually helped, I wrote an answer key for three fake syllabi in different styles (a schedule table, paragraphs, and terse lines like &lt;code&gt;HW 9: 11/6&lt;/code&gt;) plus a small script that scores a run. Good thing I did, because my first round of checks made it worse: 4 wrong items instead of 1. I'd have shipped that without the answer key.&lt;/p&gt;

&lt;p&gt;Real numbers from my laptop (i7-1355U, 16 GB RAM, running on the CPU because my MX550 only has 2 GB):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Version&lt;/th&gt;
&lt;th&gt;Found (of 20)&lt;/th&gt;
&lt;th&gt;Wrong or extra&lt;/th&gt;
&lt;th&gt;Time for 3 files&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Model only (&lt;code&gt;--raw&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;20&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;td&gt;about 6 min&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;First version of the checks&lt;/td&gt;
&lt;td&gt;20&lt;/td&gt;
&lt;td&gt;4&lt;/td&gt;
&lt;td&gt;about 8 to 11 min&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Final&lt;/td&gt;
&lt;td&gt;20&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;about 7 to 9 min&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Two runs in a row give the same output. 14 of the 20 calendar events get a real time. The other 6 just don't have a time written in the syllabus.&lt;/p&gt;

&lt;p&gt;Yes, it's slow. A first read is about 2 to 3 minutes per syllabus. But results are cached per file, so after that it only re-reads syllabi that changed, and there's a &lt;code&gt;schedule.py&lt;/code&gt; that can run it every morning so you never sit and wait. Seemed like the right amount of lazy for Sloth Day.&lt;/p&gt;

&lt;h3&gt;
  
  
  I also tried Gemma
&lt;/h3&gt;

&lt;p&gt;Since swapping models is one flag, I ran the same test with Google's Gemma 3 4B (&lt;code&gt;--model gemma3:4b&lt;/code&gt;), also local through Ollama:&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;Qwen 2.5 7B&lt;/th&gt;
&lt;th&gt;Gemma 3 4B&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Real deadlines found (of 20)&lt;/td&gt;
&lt;td&gt;20&lt;/td&gt;
&lt;td&gt;20&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Wrong items&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Time for 3 files&lt;/td&gt;
&lt;td&gt;about 7 to 9 min&lt;/td&gt;
&lt;td&gt;about 9 min&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Gemma found every real deadline and wrote slightly tidier titles, but it added a second "Research poster draft" on Nov 26, which is actually the Thanksgiving no-class row of the table. My checks let that one through. It wasn't any faster either, since my CPU is the bottleneck, not the model size. So Qwen stays the default, and Gemma is there if you want to try it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Syllabi have names, emails and office locations in them, and sometimes your own details too. With a local model, none of that leaves my laptop. It works offline, and it's free to run every day for the rest of the term.&lt;/p&gt;

&lt;p&gt;Being able to control the setup (prompt, schema, temperature) is what made the checking idea possible, and the answer key only works because the output is repeatable. Swapping models is one flag: &lt;code&gt;--model llama3.1:8b&lt;/code&gt;, or something smaller if you only have 8 GB.&lt;/p&gt;

&lt;p&gt;I didn't compare it against a hosted model, so I can't say it's more accurate. What I can say is that a 7B model on a normal laptop, with a few cheap checks around it, was accurate enough on my tests, and it stayed private, offline and free.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Lazy Payoff
&lt;/h2&gt;

&lt;p&gt;The first run takes about 7 to 9 minutes for three syllabi, but that's the laptop's time, not mine. After that, a syllabus that hasn't changed is skipped, so a daily run only reads whatever is new. In exchange I get a plan, a calendar file with a reminder on each start-by day, and a two-line morning note, without typing 20 dates into a calendar.&lt;/p&gt;

&lt;p&gt;I've only tested it on the three sample syllabi so far, so I can't give you a real-term number yet. Real syllabi are messier, which is why the README says to check the plan against the original.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;p&gt;Gemma: the planner runs Gemma 3 4B locally through Ollama (&lt;code&gt;--model gemma3:4b&lt;/code&gt;), and I tested it against Qwen on the same answer key (table above).&lt;/p&gt;

</description>
      <category>hf26challenge</category>
      <category>devchallenge</category>
      <category>gemma</category>
      <category>ai</category>
    </item>
  </channel>
</rss>
